COMPLEX HADAMARD MATRICES AND APPLICATIONS
TEO BANICA
Abstract. A complex Hadamard matrix is a square matrix H ∈ MN (C) whose entries are on the unit circle, |Hij| = 1, and whose rows and pairwise orthogonal. The main ij 2πi/N example is the Fourier matrix, FN = (w ) with w = e . We discuss here the basic theory of such matrices, with emphasis on geometric and analytic aspects.
CONTENTS
Introduction 1 1. Hadamard matrices 5 2. Complex matrices 21 3. Roots of unity 37 4. Geometry, defect 53 5. Special matrices 69 6. Circulant matrices 85 7. Bistochastic matrices 101 8. Glow computations 117 9. Norm maximizers 133 10. Quantum groups 149 11. Subfactor theory 165 12. Fourier models 181 References 197
Introduction
A complex Hadamard matrix is a square matrix H ∈ MN (C) whose entries belong to the unit circle in the complex plane, Hij ∈ T, and whose rows are pairwise orthogonal N P with respect to the usual scalar product of C , given by < x, y >= i xiy¯i.
2010 Mathematics Subject Classification. 15B10. Key words and phrases. Hadamard matrix, Fourier matrix. 1 2 TEO BANICA √ The orthogonality condition tells us that the rescaled matrix U = H/ N must be unitary. Thus, these matrices form a real algebraic manifold, given by: √ XN = MN (T) ∩ NUN ij 2πi/N The basic example is the Fourier matrix, FN = (w ) with w = e . In standard matrix form, and with indices i, j = 0, 1,...,N − 1, this matrix is as follows: 1 1 1 ... 1 1 w w2 . . . wN−1 1 w2 w4 . . . w2(N−1) FN = . . . . . . . . 1 wN−1 w(2N−1) . . . w(N−1)2 More generally, we have as example the Fourier coupling of any finite abelian group G, regarded via the isomorphism G ' Gb as a square matrix, FG ∈ MG(C): F =< i, j > G i∈G,j∈Gb
Observe that for the cyclic group G = ZN we obtain in this way the above standard Fourier matrix FN . In general, we obtain a tensor product of Fourier matrices FN .
There are many other examples of such matrices, for the most coming from various combinatorial constructions, basically involving design theory, and roots of unity. In addition, there are several deformation procedures for such matrices, leading to some more complicated constructions as well, or real algebraic geometry flavor.
In general, the complex Hadamard matrices can be thought of as being “generalized Fourier matrices”, of somewhat exotic type. Due to their generalized Fourier nature, these matrices appear in a wide array of questions in mathematics and physics:
1. Operator algebras. One important concept in the theory of von Neumann algebras is that of maximal abelian subalgebra (MASA). In the finite case, where the algebra has a trace, one can talk about pairs of orthogonal MASA. In the simplest case, of the matrix ∗ algebra MN (C), the orthogonal MASA are, up to conjugation, A = ∆,B = H∆H , where ∆ ⊂ MN (C) are the diagonal matrices, and H ∈ MN (C) is Hadamard.
2. Subfactor theory. Along the same lines, but at a more advanced level, associated ∗ to any Hadamard matrix H ∈ MN (C) is the square diagram C ⊂ ∆,H∆H ⊂ MN (C) formed by the associated MASA, which is a commuting square in the sense of subfactor theory. The Jones basic construction produces, out of this diagram, an index N subfactor of the Murray-von Neumann factor R, whose computation a key problem. HADAMARD MATRICES 3
3. Quantum groups. Associated to any complex Hadamard matrix H ∈ MN (C) is a + certain quantum permutation group G ⊂ SN , obtained by factorizing the flat representa- + tion π : C(SN ) → MN (C) associated to H. As a basic example here, the Fourier matrix FG produces in this way the group G itself. In general, the above-mentioned subfactor can be recovered from G, whose computation is a key problem.
4. Lattice models. According to the work of Jones, the combinatorics of the subfactor associated to an Hadamard matrix H ∈ MN (C), which by the above can be recovered + from the representation theory of the associated quantum permutation group G ⊂ SN , can be thought of as being the combinatorics of a “spin model”, in the context of link invariants, or of statistical mechanics, in an abstract, mathematical sense.
From a more applied point of view, the Hadamard matrices can be used in order to construct mutually unbiased bases (MUB) and other useful objects, which can help in connection with quantum information theory, and other quantum physics questions.
All this is quite recent, basically going back to the 00s. Regarding the known facts about the Hadamard matrices, most of them are in fact of purely mathematical nature. There are indeed many techniques that can be applied, leading to various results:
1. Algebra. In the real case, H ∈ MN (±1), the study of such matrices goes back to the beginning of the 20th century, and is quite advanced. The main problems, however, namely the Hadamard conjecture (HC) and the circulant Hadamard conjecture (CHC) are not solved yet, with no efficient idea of approach in sight. Part of the real matrix techniques apply quite well to the root of unity case, H ∈ MN (Zs), with s < ∞.
2. Geometry. As already explained above,√ the N × N complex Hadamard matrices form a real algebraic manifold, XN = MN (T) ∩ NUN . This manifold is highly singular, but several interesting geometric results about it have been obtained, notably about the general structure of the singularity at a given point H ∈ XN , about the neighborhood of the Fourier matrices FG, and about the various isolated points as well.
3. Analysis. One interesting point of view on the Hadamard matrices, real√ or complex, comes from the fact that these are precisely the rescaled versions, H = NU, of the P matrices which maximize the 1-norm ||U||1 = ij |Uij| on ON ,UN respectively. When looking more generally at the local maximizers of the 1-norm, one is led into a notion of “almost Hadamard matrices”, having interesting algebraic and analytic aspects.
4. Probability. Another speculative approach, this time probabilistic, is by playing a Gale-Berlekamp type game with the matrix, in the hope that the invariants which are obtained in this way are related to the various geometric and quantum algebraic invariants, 4 TEO BANICA which are hard to compute. All this is related to the subtle fact that any unitary matrix, and so any complex Hadamard matrix as well, can be put in bistochastic form. Our aim here is to survey this material, theory and applications. Organizing all this is not easy, and we have chosen an algebra/geometry/analysis/physics lineup for our presentation, vaguely coming from the amount of background which is needed. The present text is organized in 4 parts, as follows:
(1) Sections 1-3 contain basic definitions and various algebraic results. (2) Sections 4-6 deal with differential and algebraic geometric aspects. (3) Sections 7-9 are concerned with various analytic considerations. (4) Sections 10-12 deal with various mathematical physics aspects.
There are of course many aspects of the theory which are missing from our presentation, but we will provide of course some information here, comments and references.
Acknowledgements. I would like to thank Vaughan Jones for suggesting me, back to a discussion that we had in 1997, when we first met, to look at vertex models, and related topics. Stepping into bare Hadamard matrices is quite an experience, and very inspiring was the work of Uffe Haagerup on the subject, and his papers [32], [51], [52]. The present text is heavily based on a number of research papers on the subject that I wrote or co-signed, mostly during 2005–2015, and I would like to thank my cowork- ers Julien Bichon, BenoˆıtCollins, Ion Nechita, Remus Nicoar˘a,Duygu Ozteke,¨ Lorenzo Pittau, Jean-Marc Schlenker, Adam Skalski and Karol Zyczkowski.˙ Finally, many thanks go to my cats, for advice with hunting techniques, martial arts, and more. When doing linear algebra, all this knowledge is very useful. HADAMARD MATRICES 5
1. Hadamard matrices We are interested here in the complex Hadamard matrices, but we will start with some beautiful pure mathematics, regarding the real case. The definition that we need, going back to 19th century work of Sylvester [85], is as follows: Definition 1.1. An Hadamard matrix is a square binary matrix,
H ∈ MN (±1) whose rows are pairwise orthogonal, with respect to the scalar product on RN . As a first observation, we do not really need real numbers in order to talk about the Hadamard matrices, because the orthogonality condition tells us that, when comparing two rows, the number of matchings should equal the number of mismatchings. Thus, we can replace if we want the 1, −1 entries of our matrix by any two symbols, of our choice. Here is an example of an Hadamard matrix, with this convention: ♥ ♥ ♥ ♥ ♥ ♣ ♥ ♣ ♥ ♥ ♣ ♣ ♥ ♣ ♣ ♥ However, it is probably better to run away from this, and use real numbers instead, as in Definition 1.1, with the idea in mind of connecting the Hadamard matrices to the foundations of modern mathematics, namely Calculus 1 and Calculus 2. So, getting back now to the real numbers, here is a first result: Proposition 1.2. The set of the N × N Hadamard matrices is √ YN = MN (±1) ∩ NON where ON is the orthogonal group, the intersection being taken inside MN (R). √ Proof. Let H ∈ MN (±1). Since the rows of the rescaled matrix U = H/ N have norm 1, with respect to the usual scalar product on RN , we conclude that H is Hadamard precisely when U belongs to the orthogonal group ON , and so when H ∈ YN , as claimed. As an interesting consequence of the above result, which is not exactly obvious when using the design theory approach, we have the following result:
Proposition 1.3. Let H ∈ MN (±1) be an Hadamard matrix. (1) The columns of H must be pairwise orthogonal. t (2) The transpose matrix H ∈ MN (±1) is Hadamard as well.
Proof. Since the orthogonal group ON is stable under transposition, so is the set YN constructed in Proposition 1.2, and this gives both the assertions. 6 TEO BANICA
Let us study now the examples. There are many such matrices, and in order to cut a bit from the complexity, we can use the following notions:
Definition 1.4. Two Hadamard matrices are called equivalent, and we write H ∼ K, when it is possible to pass from H to K via the following operations: (1) Permuting the rows, or the columns. (2) Multiplying the rows or columns by −1. Also, we say that H is dephased when its first row and column consist of 1 entries.
Observe that we do not include the transposition operation H → Ht in our list of allowed operations. This is because Proposition 1.3 above, while looking quite elementary, rests however on a deep linear algebra fact, namely that the transpose of an orthogonal matrix is orthogonal as well, and this can produce complications later on. Regarding the equivalence, there is of course a certain group G acting there, made of N two copies of SN , one for the rows and one for the columns, and of two copies of Z2 , once again one for the rows, and one for the columns. The equivalence classes of the Hadamard matrices are then the orbits of the action G y YN . It is possible to be a bit more explicit here, with a formula for G and so on, but we will not need this. As for the dephasing, here the terminology comes from physics, or rather from the complex Hadamard matrices. Indeed, when regarding H ∈ MN (±1) as a complex matrix, H ∈ MN (T), the −1 entries have “phases”, equal to π, and assuming that H is dephased means to assume that we have no phases, on the first row and the first column. Observe that, up to the equivalence relation, any Hadamard matrix H ∈ MN (±1) can be put in dephased form. Moreover, the dephasing operation is unique, if we use only the operations (2) in Definition 1.4, namely row and column multiplications by −1. With these notions in hand, we can formulate our first classification result:
Proposition 1.5. There is only one Hadamard matrix at N = 2, namely
1 1 W = 2 1 −1 up to the above equivalence relation for such matrices.
Proof. The matrix in the statement W2, called Walsh matrix, is clearly Hadamard. Con- versely, given H ∈ MN (±1) Hadamard, we can dephase it, as follows:
a b 1 1 1 1 → → c d ac bd 1 abcd
Now since the dephasing operation preserves the class of the Hadamard matrices, we must have abcd = −1, and so we obtain by dephasing the matrix W2. HADAMARD MATRICES 7
At N = 3 we cannot have examples, due to the orthogonality condition, which forces N to be even. At N = 4 now, we have several examples, as for instance: 1 1 1 1 1 −1 1 −1 W4 = 1 1 −1 −1 1 −1 −1 1 This matrix is a particular case of the following construction:
Proposition 1.6. If H ∈ MM (±1) and K ∈ MN (±1) are Hadamard matrices, then so is their tensor product, constructed in double index notation as follows:
H ⊗ K ∈ MMN (±1) , (H ⊗ K)ia,jb = HijKab ⊗n n In particular the Walsh matrices, WN = W2 with N = 2 , are all Hadamard.
Proof. The matrix in the statement H ⊗ K has indeed ±1 entries, and its rows Ria are pairwise orthogonal, as shown by the following computation: X < Ria,Rkc > = HijKab · HkjKcb jb X X = HijHkj KabKcb j b
= MNδikδac
As for the second assertion, this follows from this, W2 being Hadamard. Before going further, we should perhaps clarify a bit our tensor product notations. In order to write H ∈ MN (±1) the indices of H must belong to {1,...,N}, or at least to an ordered set {I1,...,IN }. But with double indices we are indeed in this latter situation, because we can use the lexicographic order on these indices. To be more precise, by using the lexicographic order on the double indices, we have the following formula:
H11K...H1M K . . H ⊗ K = . . HM1K...HMM K
As an example, by tensoring W2 with itself, we obtain the above matrix W4. Getting back now to our classification work, here is the result at N = 4: Proposition 1.7. There is only one Hadamard matrix at N = 4, namely
W4 = W2 ⊗ W2 up to the standard equivalence relation for such matrices. 8 TEO BANICA
Proof. Consider an Hadamard matrix H ∈ M4(±1), assumed to be dephased: 1 1 1 1 1 a b c H = 1 d e f 1 g h i By orthogonality of the first 2 rows we must have {a, b, c} = {−1, −1, 1}, and so by permuting the last 3 columns, we can further assume that our matrix is as follows: 1 1 1 1 1 −1 1 −1 H = 1 m n o 1 p q r By orthogonality of the first 2 columns we must have {m, p} = {−1, 1}, and so by permuting the last 2 rows, we can further assume that our matrix is as follows: 1 1 1 1 1 −1 1 −1 H = 1 1 x y 1 −1 z t But this gives the result, because from the orthogonality of the rows we obtain x = y = −1, and then, with these values of x, y plugged in, from the orthogonality of the columns we obtain z = −1, t = 1. Thus, up to equivalence we have H = W4, as claimed. The case N = 5 is excluded, because the orthogonality condition forces N ∈ 2N. The point now is that the case N = 6 is excluded as well, because we have:
Proposition 1.8. The size of an Hadamard matrix must be N ∈ {2} ∪ 4N. Proof. By permuting the rows and columns or by multiplying them by −1, as to rearrange the first 3 rows, we can always assume that our matrix looks as follows: 1 ...... 1 1 ...... 1 1 ...... 1 1 ...... 1 1 ...... 1 1 ...... 1 −1 ... − 1 −1 ... − 1 H = 1 ...... 1 −1 ... − 1 1 ...... 1 −1 ... − 1 ...... | {z } | {z } | {z } | {z } x y z t Now if we denote by x, y, z, t the sizes of the 4 block columns, as indicated, the orthog- onality conditions between the first 3 rows give the following system of equations: (1 ⊥ 2) : x + y = z + t (1 ⊥ 3) : x + z = y + t (2 ⊥ 3) : x + t = y + z HADAMARD MATRICES 9
The solution of this system being x = y = z = t, we conclude that the size of our matrix N = x + y + z + t must be a multiple of 4, as claimed. As a consequence, we are led to the study of the Hamadard matrices at: N = 8, 12, 16, 20, 24,... This study can be done either abstractly, via various algebraic methods, or with a computer, and this leads to the conclusion that the number of Hadamard matrices of size N ∈ 4N grows with N, and this in a rather exponential fashion. In particular, we are led in this way into the following statement: Conjecture 1.9 (Hadamard Conjecture (HC)). There is at least one Hadamard matrix
H ∈ MN (±1) for any integer N ∈ 4N. This conjecture, going back to the 19th century, is probably one of the most beautiful statements in combinatorics, linear algebra, and mathematics in general. Quite remark- ably, the numeric verification so far goes up to the number of the beast: N = 666 Our purpose now will be that of gathering some evidence for this conjecture. At N = 8 we have the Walsh matrix W8. Thus, the next existence problem comes at N = 12. And here, we can use the following key construction, due to Paley [75]:
r Theorem 1.10. Let q = p be an odd prime power, define χ : Fq → {−1, 0, 1} by χ(0) = 0, χ(a) = 1 if a = b2 for some b 6= 0, and χ(a) = −1 otherwise, and finally set Qab = χ(a − b). We have then constructions of Hadamard matrices, as follows: (1) Paley 1: if q = 3(4) we have a matrix of size N = q + 1, as follows: 0 1 ... 1 −1 1 P = 1 + . N . Q −1 (2) Paley 2: if q = 1(4) we have a matrix of size N = 2q + 2, as follows: 0 1 ... 1 1 2 1 −1 1 1 P = . : 0 → , ±1 → ± N . Q −1 −1 1 −1 1 These matrices are skew-symmetric (H + Ht = 2), respectively symmetric (H = Ht). 10 TEO BANICA
Proof. In order to simplify the presentation, we will denote by 1 all the identity matrices, of any size, and by I all the rectangular all-one matrices, of any size as well. It is elementary to check that the matrix Qab = χ(a − b) has the following properties: t QQ = q1 − I ,QI = IQ = 0 In addition, we have the following formulae, which are elementary as well, coming from the fact that −1 is a square in Fq precisely when q = 1(4): q = 1(4) =⇒ Q = Qt q = 3(4) =⇒ Q = −Qt With these observations in hand, the proof goes as follows: (1) With our conventions for the symbols 1 and I, explained above, the matrix in the statement is as follows: 1 P 1 = I N −I 1 + Q With this formula in hand, the Hadamard matrix condition follows from: 1 1 − P 1 (P 1 )t = I I N N −I 1 + Q I 1 − Q N 0 = 0 I + 1 − Q2 N 0 = 0 N (2) If we denote by G, F the matrices in the statement, which replace respectively the 0, 1 entries, then we have the following formula for our matrix: 0 P 2 = I ⊗ F + 1 ⊗ G N I Q With this formula in hand, the Hadamard matrix condition follows from: 0 2 1 0 0 (P 2 )2 = I ⊗ F 2 + ⊗ G2 + I ⊗ (FG + GF ) N I Q 0 1 I Q q 0 1 0 0 = ⊗ 2 + ⊗ 2 + I ⊗ 0 0 q 0 1 I Q N 0 = 0 N t Finally, the last assertion is clear, from the above formulae relating Q, Q . These constructions allow us to get well beyond the Walsh matrix level, and we have the following result: HADAMARD MATRICES 11
Theorem 1.11. The HC is verified at least up to N = 88, as follows: (1) At N = 4, 8, 16, 32, 64 we have Walsh matrices. (2) At N = 12, 20, 24, 28, 44, 48, 60, 68, 72, 80, 84, 88 we have Paley 1 matrices. (3) At N = 36, 52, 76 we have Paley 2 matrices. (4) At N = 40, 56 we have Paley 1 matrices tensored with W2. However, at N = 92 these constructions (Walsh, Paley, tensoring) don’t work. Proof. First of all, the numbers in (1-4) are indeed all the multiples of 4, up to 88. As for the various assertions, the proof here goes as follows: (1) This is clear. (2) Since N − 1 takes the values q = 11, 19, 23, 27, 43, 47, 59, 67, 71, 79, 83, 87, all prime powers, we can indeed apply the Paley 1 construction, in all these cases. (3) Since N = 4(8) here, and N/2 − 1 takes the values q = 17, 25, 37, all prime powers, we can indeed apply the Paley 2 construction, in these cases. 1 1 (4) At N = 40 we have indeed P20 ⊗ W2, and at N = 56 we have P28 ⊗ W2. Finally, we have 92 − 1 = 7 × 13, so the Paley 1 construction does not work, and 92/2 = 46, so the Paley 2 construction, or tensoring with W2, does not work either. At N = 92 the situation is considerably more complicated, and we have:
Theorem 1.12. Assuming that A, B, C, D ∈ MK (±1) are circulant, symmetric, pairwise commute and satisfy A2 + B2 + C2 + D2 = 4K, the following 4K × 4K matrix ABCD −BA −DC H = −CDA −B −D −CBA is Hadamard, called of Williamson type. Moreover, such a matrix exists at K = 23.
Proof. The matrix H can be written as follows, where 1, i, j, k ∈ M4(0, 1), called the quaternion units, are the matrices describing the positions of the A, B, C, D entries: H = A ⊗ 1 + B ⊗ i + C ⊗ j + D ⊗ k Assuming now that A, B, C, D are symmetric, we have: HHt = (A ⊗ 1 + B ⊗ i + C ⊗ j + D ⊗ k)(A ⊗ 1 − B ⊗ i − C ⊗ j − D ⊗ k) = (A2 + B2 + C2 + D2) ⊗ 1 − ([A, B] − [C,D]) ⊗ i −([A, C] − [B,D]) ⊗ j − ([A, D] − [B,C]) ⊗ k Thus, if we further assume that A, B, C, D pairwise commute, and satisfy the condition A2 + B2 + C2 + D2 = 4K, we obtain indeed an Hadamard matrix. In general, finding such matrices is a difficult task, and this is where Williamson’s extra assumption that A, B, C, D should be taken circulant comes from. 12 TEO BANICA
Regarding now the K = 23 construction, which produces an Hadamard matrix of order N = 92, this comes via a computer search. We refer here to [24], [97]. Things get even worse at higher values of N, where more and more complicated con- structions are needed. The whole subject is quite technical, and, as already mentioned, human knowledge here stops so far at N = 666. See [1], [41], [43], [56], [64], [84]. As a conceptual finding on this subject, however, we have the recent theory of the cocyclic Hadamard matrices. The basic definition here is as follows:
Definition 1.13. A cocycle on a finite group G is a matrix H ∈ MG(±1) satisfying:
H11 = 1 ,HghHgh,k = Hg,hkHhk If the rows of H are pairwise orthogonal, we say that H is a cocyclic Hadamard matrix. Here the definition of the cocycles is the usual one, with the equations coming from the fact that F = Z2 × G must be a group, with multiplication as follows:
(u, g)(v, h) = (Hgh · uv, gh)
As a basic example, the Walsh matrix H = W2n is cocyclic, coming from the group n
One potential way of getting away from these questions is that of looking at various special classes of Hadamard matrices. However, this is not really the case, because passed a few trivialities, the existence of special Hadamard matrices is generally subject to an improvement of the HC, as in the cocyclic case, or to difficult non-existence questions. Illustrating and quite famous here is the situation in the circulant case. Given a vector N γ ∈ (±1) , one can ask whether the matrix H ∈ MN (±1) defined by Hij = γj−i is Hadamard or not. Here is a solution to the problem: −1 1 1 1 1 −1 1 1 K4 = 1 1 −1 1 1 1 1 −1
4 P More generally, any vector γ ∈ (±1) satisfying γi = ±1 is a solution to the problem. The following conjecture, going back to [84], states that there are no other solutions: HADAMARD MATRICES 13
Conjecture 1.15 (Circulant Hadamard Conjecture (CHC)). There is no circulant Ha- damard matrix of size N × N, for any N 6= 4. The fact that such a simple-looking problem is still open might seem quite surprising. Indeed, if we denote by S ⊂ {1,...,N} the set of positions of the −1 entries of γ, the Hadamard matrix condition is simply |S ∩ (S + k)| = |S| − N/4, for any k 6= 0, taken modulo N. Thus, the above conjecture simply states that at N 6= 4, such a set S cannot exist. Let us record here this latter statement, originally due to Ryser [83]: Conjecture 1.16 (Ryser Conjecture). Given an integer N > 4, there is no set S ⊂ {1,...,N} satisfying the condition |S ∩ (S + k)| = |S| − N/4 for any k 6= 0, taken modulo N. There has been a lot of work on this conjecture, starting with [83]. However, as it was the case with the HC, all this leads to complicated combinatorics, design theory, algebra and number theory, and so on, and there is no serious idea here, at least so far.
All this might seem a bit depressing, but there are at least two potential exits from the combinatorial and algebraic theory of Hadamard matrices, namely: (1) Do analysis. There are many things that can be done here, starting with the Hadamard determinant bound [53], and we will discuss this below. Whether all this can help or not in relation with the HC and CHC remains to be seen, but at least we’ll have some fun, and do some interesting mathematics. (2) Do physics. When allowing the entries of H to be complex numbers, both the HC ij 2πi/N and the CHC dissapear, because the Fourier matrix FN = (w ) with w = e is an example of such matrix at any N ∈ N, which in addition can be put in circulant form. We will discuss this later, starting from section 2 below. So, let us step now into analytic questions. The first result here, found in 1893 by Hadamard [53], about 25 years after Sylvester’s 1867 founding paper [85], and which actually led to such matrices being called Hadamard, is as follows:
Theorem 1.17. Given a matrix H ∈ MN (±1), we have | det(H)| ≤ N N/2 with equality precisely when H is Hadamard. Proof. We use here the fact, which often tends to be forgotten, that the determinant of a system of N vectors in RN is the signed volume of the associated parallelepiped:
det(H1,...,HN ) = ±vol < H1,...,HN > This is actually the definition of the determinant, in case you have forgotten the basics (!), with the need for the sign coming for having good additivity properties. 14 TEO BANICA
In the case where our vectors take their entries in ±1, we therefore have the following inequality, with equality precisely when our vectors are pairwise orthogonal: √ N | det(H1,...,HN )| ≤ ||H1|| × ... × ||HN || = ( N) Thus, we have obtained the result, straight from the definition of det. The above result is quite interesting, philosophically speaking. Let us recall indeed from Proposition 1.2 that the set formed by the N × N Hadamard matrices is: √ YN = MN (±1) ∩ NON
Thus, what we have in Theorem 1.17 is an analytic method for locating YN inside MN (±1). This suggests doing many geometric and analytic things, as for instance looking at the maximizers of | det(H)| at values N ∈ N which are not multiples of 4. These latter matrices are called “quasi-Hadamard”, and we refer here to [76]. √ From a “dual” point of view, the question of locating YN inside NON , once again via analytic methods, makes sense as well. The result here, from [14], is as follows:
Theorem 1.18. Given a matrix U ∈ ON we have √ ||U||1 ≤ N N √ with equality precisely when H = U/ N is Hadamard.
Proof. We have indeed the following estimate, valid for any U ∈ ON : !1/2 √ X X 2 ||U||1 = |Uij| ≤ N |Uij| = N N ij ij √ The equality√ case holds when |Uij| = N for any i, j, which amounts in saying that H = U/ N must satisfy H ∈ MN (±1), and so that H must be Hadamard. As a first comment here, the above Cauchy-Schwarz estimate can be improved with a H¨olderestimate, the conclusion being that the rescaled Hadamard matrices maximize the p-norm on ON at any p ∈ [1, 2), and minimize it at any p ∈ (2, ∞]. We will discuss this in section 9 below, with full details, directly in the complex case. As it was the case with the Hadamard determinant bound, all this suggests doing some further geometry and analysis, this time on the Lie group ON , notably with a notion of “almost Hadamard matrix” at stake. We will be back to this in section 9 below.
Let us discuss now, once again as an introduction to analytic topics, yet another such result. We recall that a matrix H ∈ MN (R) is called row-stochastic when the sums on the rows are all equal, column-stochastic when the same is true for columns, and bistochastic when this is true for both rows and columns, the common sum being the same. With this terminology, we have the following well-known result: HADAMARD MATRICES 15
Theorem 1.19. For an Hadamard matrix H ∈ MN (±1), the excess, X E(H) = Hij ij √ satisfies |E(H)| ≤ N N, with equality if and only if H is bistochastic.
N Proof. In terms of the all-one vector ξ = (1)i ∈ R , we have: X X X E(H) = Hij = Hijξjξi = (Hξ)iξi =< Hξ, ξ > ij ij i √ Now by using the Cauchy-Schwarz inequality, along with the fact that U = H/ N is orthogonal, and hence of norm 1, we obtain, as claimed: √ |E(H)| ≤ ||Hξ|| · ||ξ|| ≤ ||H|| · ||ξ||2 = N N
Regarding now the equality case, this requires the vectors√Hξ, ξ to be proportional, and so our matrix H to be row-stochastic. But since U = H/ N is orthogonal, Hξ ∼ ξ t is equivalent to H ξ ∼ ξ, and so H must be bistochastic, as claimed. There are many known interesting results on the bistochastic Hadamard matrices, and we will be back to this in section 7 below, directly in the complex setting.
One interesting question, that we would like to discuss now, is that of computing the law of the excess over the equivalence class of H. Let us start with:
Definition 1.20. The glow of H ∈ MN (±1) is the probability measure µ ∈ P(Z) describ- P ing the distribution of the excess, E = ij Hij, over the equivalence class of H. Since the excess is invariant under permutations of rows and columns, we can restrict the attention to the matrices He ' H obtained by switching signs on rows and columns. N N More precisely, let (a, b) ∈ Z2 × Z2 , and consider the following matrix:
Heij = aibjHij
N N We can regard the sum of entries of He as a random variable, over the group Z2 × Z2 , and we have the following equivalent description of the glow:
N N Proposition 1.21. Given a matrix H ∈ MN (±1), if we define ϕ : Z2 × Z2 → Z by X ϕ(a, b) = aibjHij ij then the glow µ is the probability measure on Z given by µ({k}) = P (ϕ = k). 16 TEO BANICA
Proof. The function ϕ in the statement can indeed be regarded as a random variable over N N the group Z2 × Z2 , with this latter group being endowed with its uniform probability measure P . The distribution µ of this variable ϕ is then given by: 1 n o µ({k}) = # (a, b) ∈ N × N ϕ(a, b) = k 4N Z2 Z2 By the above discussion, this distribution is exactly the glow. The terminology in Definition 1.20 comes from the following picture. Assume that we have a square city, with N horizontal streets and N vertical streets, and with street lights at each crossroads. When evening comes the lights are switched on at the positions (i, j) where Hij = 1, and then, all night long, they are randomly switched on and off, with the help of 2N master switches, one at the end of each street: → ♦ ♦ ♦ ♦ → ♦ × ♦ × → ♦ ♦ × × → ♦ × × ♦
↑ ↑ ↑ ↑ With this picture in mind, µ describes indeed the glow of the city. At a more advanced level now, all this is related to the Gale-Berlekamp game [50], [82], and this is where our main motivation for studying it comes from. In order to compute the glow, it is useful to have in mind the following picture:
b1 . . . bN ↓ ↓ (a1) → H11 ...H1N ⇒ S1 ...... (aN ) → HN1 ...HNN ⇒ SN Here the columns of H have been multiplied by the entries of the horizontal switching vector b, the resulting sums on rows are denoted S1,...,SN , and the vertical switching vector a still has to act on these sums, and produce the glow component at b. With this picture in mind, we first have the following result, from [9]:
Proposition 1.22. The glow of a matrix H ∈ MN (±1) is given by 1 X µ = β (c ) ∗ ... ∗ β (c ) 2N 1 1 N N N b∈Z2 ∗c δr+δ−r where βr(c) = 2 , and cr = #{r ∈ |S1|,..., |SN |}, with S = Hb. HADAMARD MATRICES 17
Proof. We use the interpretation of the glow explained above. So, consider the decompo- sition of the glow over b components: 1 X µ = µ 2N b N b∈Z2
With the notation S = Hb, the numbers S1,...,SN are the sums on the rows of the matrix Heij = Hijaibj. Thus the glow components are given by:
µb = law (±S1 ± S2 ... ± SN ) By permuting now the sums on the right, we have the following formula: µb = law ±0 ... ± 0 ±1 ... ± 1 ...... ±N... ± N | {z } | {z } | {z } c0 c1 cN Now since the ± variables each follow a Bernoulli law, and these Bernoulli laws are independent, we obtain a convolution product as in the statement. We will need the following elementary fact:
Proposition 1.23. Let H ∈ MN (±1) be an Hadamard matrix of order N ≥ 4.
(1) The sums of entries on rows S1,...,SN are even, and equal modulo 4. (2) If the sums on the rows S1,...,SN are all 0 modulo 4, then the number of rows whose sum is 4 modulo 8 is odd for N = 4(8), and even for N = 0(8). Proof. (1) Let us pick two rows of our matrix, and then permute the columns such that these two rows look as follows: 1 ...... 1 1 ...... 1 −1 ... − 1 −1 ... − 1 1 ...... 1 −1 ... − 1 1 ...... 1 −1 ... − 1 | {z } | {z } | {z } | {z } a b c d N We have a + b + c + d = N, and by orthogonality a + d = b + c, so a + d = b + c = 2 . Now since N/2 is even, we conclude that b = c(2), and this gives the result. (2) In the case where H is “row-dephased”, in the sense that its first row consists of 1 entries only, the row sums are N, 0,..., 0, and so the result holds. In general now, by permuting the columns we can assume that our matrix looks as follows: 1 ...... 1 −1 ... − 1 . . H = . . |{z} |{z} x y
We have x + y = N = 0(4), and since the first row sum S1 = x − y is by assumption 0 modulo 4, we conclude that x, y are even. In particular, since y is even, the passage from H to its row-dephased version He can be done via y/2 double sign switches. Now, in view of the above, it is enough to prove that the conclusion in the statement is stable under a double sign switch. So, let H ∈ MN (±1) be Hadamard, and let us perform 18 TEO BANICA to it a double sign switch, say on the first two columns. Depending on the values of the entries on these first two columns, the total sums on the rows change as follows: + + ...... : S → S − 4 + − ...... : S → S − + ...... : S → S − − ...... : S → S + 4 We can see that the changes modulo 8 of the row sum S occur precisely in the first and in the fourth case. But, since the first two columns of our matrix H ∈ MN (±1) are orthogonal, the total number of these cases is even, and this finishes the proof. Observe that Proposition 1.22 and Proposition 1.23 (1) show that the glow of an Ha- damard matrix of order N ≥ 4 is supported by 4Z. With this in hand, we have:
Theorem 1.24. Let H ∈ MN (±1) be an Hadamard matrix of order N ≥ 4, and denote by µeven, µodd the mass one-rescaled restrictions of µ ∈ P(4Z) to 8Z, 8Z + 4. 3 even 1 odd (1) At N = 0(8) we have µ = 4 µ + 4 µ . 1 even 3 odd (2) At N = 4(8) we have µ = 4 µ + 4 µ . Proof. We use the glow decomposition over b components, from Proposition 1.22: 1 X µ = µ 2N b N b∈Z2 The idea is that the decomposition formula in the statement will occur over averages N−1 of the following type, over truncated sign vectors c ∈ Z2 : 1 µ0 = (µ + µ ) c 2 +c −c Indeed, we know from Proposition 1.23 (1) that modulo 4, the sums on rows are either 0,..., 0 or 2,..., 2. Now since these two cases are complementary when pairing switch vectors (+c, −c), we can assume that we are in the case 0,..., 0 modulo 4. Now by looking at this sequence modulo 8, and letting x be the number of 4 components, so that the number of 0 components is N − x, we have: 1 1 (µ+c + µ−c) = law(±0 ... ± 0 ±4 ... ± 4) + law(±2 ... ± 2) 2 2 | {z } | {z } | {z } N−x x N Now by using Proposition 1.23 (2), the first summand splits 1−0 or 0−1 on 8Z, 8Z+4, depending on the class of N modulo 8. As for the second summand, since N is even this 1 1 3 1 always splits 2 − 2 on 8Z, 8Z + 4. So, by making the average we obtain either a 4 − 4 or 1 3 a 4 − 4 splitting on 8Z, 8Z + 4, depending on the class of N modulo 8, as claimed. HADAMARD MATRICES 19
Various computer simulations suggest that the measures µeven, µodd don’t have further general algebraic properties. Analytically speaking now, we have:
Theorem 1.25. The glow moments of H ∈ MN (±1) are given by: Z E 2p = (2p)!! + O(N −1) N N N Z2 ×Z2 In particular the variable E/N becomes Gaussian in the N → ∞ limit.
Proof. Let Peven(r) ⊂ P (r) be the set of partitions of {1, . . . , r} having all blocks of even P size. The moments of the variable E = ij aibjHij are then given by: Z Z Z r X E = Hi1x1 ...Hirxr ai1 . . . air bx1 . . . bxr N N N N Z2 ×Z2 ix Z2 Z2 X X = Hi1x1 ...Hirxr
π,σ∈Peven(r) ker i=π,ker x=σ
Thus the moments decompose over partitions π ∈ Peven(r), with the contributions being obtained by integrating the following quantities: X X C(σ) = Hi1x1 ...Hirxr · ai1 . . . air ker x=σ i Now by M¨obiusinversion, we obtain a formula as follows: Z X Er = K(π)N |π|I(π) N × N Z2 Z2 π∈Peven(r) Here K(π) = P µ(π, σ), where µ is the M¨obiusfunction of P (r), and, with σ∈Peven(r) even N the convention that H1,...,HN ∈ Z2 are the rows of H: * + X Y 1 Y I(π) = H , 1 N ir i b∈π r∈b With this formula in hand, the first assertion follows, because the biggest elements of the lattice Peven(2p) are the (2p)!! partitions consisting of p copies of a 2-block. As for the second assertion, this follows from the moment formula, and from the fact that the glow of H ∈ MN (±1) is real, and symmetric with respect to 0. See [8]. We will be back to all this in section 8 below, in the complex matrix setting.
Finally, some interesting analytic results can be obtained by exiting the square matrix setting, and looking at the rectangular matrix case. Let us start with: 20 TEO BANICA
Definition 1.26. A partial Hadamard matrix (PHM) is a matrix
H ∈ MM×N (±1) having its rows pairwise orthogonal. These matrices are quite interesting objects, appearing in connection with various questions in combinatorics. The motivating examples are the Hadamard matrices H ∈ MN (±1), and their M × N submatrices, with M ≤ N. See [54], [58], [84], [94]. In their paper [44], de Launey and Levin were able to count these matrices, in the asymptotic limit N ∈ 4N, N → ∞. Their method is based on:
Proposition 1.27. The probability for a random H ∈ MM×N (±1) to be partial Hadamard equals the probability for a length N random walk with increments drawn from n o M E = (eie¯j)i Proof. Indeed, with T (e) = (eie¯j)i Theorem 1.28. The probability for a random H ∈ MM×N (±1) to be PHM is 2(M−1)2 PM ' q M (2πN)( 2 ) in the N ∈ 4N, N → ∞ limit. Proof. According to Proposition 1.27 above, we have: ( ) 1 X 1 X P = # ξ , . . . , ξ ∈ E ξ = 0 = δ M q(M−1)N 1 N i q(M−1)N Σξi,0 i ξ1,...,ξN ∈E M By using the Fourier inversion formula we have, with D = 2 : 1 Z i<λ,Σξi> δΣξi,0 = D e dλ (2π) [−π,π]D After many non-trivial computations, this leads to the result. See [44]. All this is extremely interesting. Let us mention as well that for the general matrices H ∈ MM×N (±1), which are not necessarily PHM, such statistics can be deduced from the work of Tau-Vu [91]. Finally, there is an extension of the notion of PHM in the complex case, and we will be back to this later on, in section 3 below. HADAMARD MATRICES 21 2. Complex matrices We have seen that the Hadamard matrices H ∈ MN (±1) are very interesting combina- torial objects. In what follows, we will be interested in their complex versions: Definition 2.1. A complex Hadamard matrix is a square complex matrix H ∈ MN (C) whose entries are on the unit circle, Hij ∈ T, and whose rows are pairwise orthogonal. Here, and in what follows, the scalar product is the usual one on CN , taken to be linear in the first variable and antilinear in the second one: X < x, y >= xiy¯i i As basic examples of complex Hamadard matrices, we have of course the real Hadamard matrices, H ∈ MN (±1). We will see that there are many other examples. Let us start by extending some basic results from the real case. First, we have: Proposition 2.2. The set of the N × N complex Hadamard matrices is the real algebraic manifold √ XN = MN (T) ∩ NUN where UN is the unitary group, the intersection being taken inside MN (C). √ Proof. Let H ∈ MN (T). Then H is Hadamard if and only if its rescaling U = H/ N belongs to the unitary group UN , and so when H ∈ YN , as claimed. The above manifold XN , while appearing by definition as an intersection of smooth manifolds, is very far from being smooth. We will be back to this, later on. As a basic consequence of the above result, we have: Proposition 2.3. Let H ∈ MN (C) be an Hadamard matrix. (1) The columns of H must be pairwise orthogonal. t ¯ ∗ (2) The matrices H , H,H ∈ MN (C) are Hadamard as well. Proof. We use the well-known fact that if a matrix is unitary, U ∈ UN , then so is its ¯ ¯ complex conjugate U = (Uij), the inversion formulae being as follows: U ∗ = U −1 ,U t = U¯ −1 t ¯ ∗ Thus the unitary group UN is stable under the operations U → U ,U → U,U → U , and it follows that the algebraic manifold XN constructed in Proposition 2.2 is stable as well under these operations. But this gives all the assertions. Let us introduce now the following equivalence notion for the complex Hadamard ma- trices, taking into account some basic operations which can be performed: 22 TEO BANICA Definition 2.4. Two complex Hadamard matrices are called equivalent, and we write H ∼ K, when it is possible to pass from H to K via the following operations: (1) Permuting the rows, or permuting the columns. (2) Multiplying the rows or columns by numbers in T. Also, we say that H is dephased when its first row and column consist of 1 entries. The same remarks as in the real case apply. For instance, we have not taken into account the results in Proposition 2.3 when formulating the above definition, because the operations H → Ht, H,H¯ ∗ are far more subtle than those in (1,2) above. At the level of the examples now, we have the following basic construction, which works at any N ∈ N, in stark contrast with what happens in the real case: ij 2πi/N Theorem 2.5. The Fourier matrix, FN = (w ) with w = e , which in standard matrix form, with indices i, j = 0, 1,...,N − 1, is as follows, 1 1 1 ... 1 1 w w2 . . . wN−1 1 w2 w4 . . . w2(N−1) FN = . . . . . . . . 1 wN−1 w(2N−1) . . . w(N−1)2 is a complex Hadamard matrix, in dephased form. Proof. By using the standard fact that the averages of complex numbers correspond to barycenters, we conclude that the scalar products between the rows of FN are: X aj −bj X (a−b)j < Ra,Rb >= w w = w = Nδab j j Thus FN is indeed a complex Hadamard matrix. As for the fact that FN is dephased, this follows from our convention i, j = 0, 1,...,N − 1, which is there for this. Thus, there is no analogue of the HC in the complex case. We will see later on, in section 6 below, that the Fourier matrix FN can be put in circulant form, so there is no analogue of the CHC either, in this setting. This is of course very good news. We should mention, however, that the HC and CHC do have some complex extensions, which are of technical nature, by restricting the attention to the Hadamard matrices formed by roots of unity of a given order. We will discuss this in section 3 below. As a first classification result now, in the complex case, we have: Proposition 2.6. The Fourier matrices F2,F3, which are given by 1 1 1 1 1 F = ,F = 1 w w2 2 1 −1 3 1 w2 w with w = e2πi/3 are the only Hadamard matrices at N = 2, 3, up to equivalence. HADAMARD MATRICES 23 Proof. The proof at N = 2 is similar to the proof of Proposition 1.5. Regarding now the case N = 3, consider an Hadamard matrix H ∈ M3(T), in dephased form: 1 1 1 H = 1 x y 1 z t The orthogonality conditions between the rows of this matrix read: (1 ⊥ 2) : x + y = −1 (1 ⊥ 3) : z + t = −1 (2 ⊥ 3) : xz¯ + yt¯= −1 Now observe that the equation p+q = −1 with p, q ∈ T tells us that the triangle having vertices at 1, p, q must be equilateral, and so that {p, q} = {w, w2}, with w = e2πi/3. By using this fact, for the first two equations, we conclude that we must have {x, y} = {w, w2} and {z, t} = {w, w2}. As for the third equation, this tells us that we must have x 6= z. Thus, our Hadamard matrix H is either the Fourier matrix F3, or is the matrix obtained from F3 by permuting the last two columns, and we are done. In order to deal now with the case N = 4, we already know, from our study in the real case, that we will need tensor products. So, let us formulate: Definition 2.7. The tensor product of complex Hadamard matrices is given, in double indices, by (H ⊗ K)ia,jb = HijKab. In other words, we have the formula H11K...H1M K . . H ⊗ K = . . HM1K...HMM K by using the lexicographic order on the double indices. In order to advance, our first task will be that of tensoring the Fourier matrices. And here, we have the following statement, refining and generalizing Theorem 2.5: Theorem 2.8. Given a finite abelian group G, with dual group Gb = {χ : G → T}, consider the Fourier coupling FG : G × Gb → T, given by (i, χ) → χ(i). (1) Via the standard isomorphism G ' Gb, this Fourier coupling can be regarded as a square matrix, FG ∈ MG(T), which is a complex Hadamard matrix. (2) In the case of the cyclic group G = ZN we obtain in this way, via the standard identification ZN = {1,...,N}, the Fourier matrix FN . (3) In general, when using a decomposition G = ZN1 × ... × ZNk , the corresponding Fourier matrix is given by FG = FN1 ⊗ ... ⊗ FNk . 24 TEO BANICA Proof. This follows indeed from some basic facts from group theory: (1) With the identification G ' Gb made our matrix is given by (FG)iχ = χ(i), and the scalar products between the rows are computed as follows: X X < Ri,Rj >= χ(i)χ(j) = χ(i − j) = |G| · δij χ χ Thus, we obtain indeed a complex Hadamard matrix. (2) This follows from the well-known and elementary fact that, via the identifications ij 2πi/N ZN = ZcN = {1,...,N}, the Fourier coupling here is (i, j) → w , with w = e . (3) We use here the following well-known formula, for the duals of products: H\× K = Hb × Kb At the level of the corresponding Fourier couplings, we obtain from this: FH×K = FH ⊗ FK Now by decomposing G into cyclic groups, as in the statement, and by using (2) for the cyclic components, we obtain the formula in the statement. As a first application of this result, we have: n Proposition 2.9. The Walsh matrix, WN with N = 2 , which is given by 1 1 ⊗n W = N 1 −1 n is the Fourier matrix of the finite abelian group KN = Z2 . Proof. We have indeed W2 = F2 = FK2 , and by taking tensor powers we obtain from this n that we have WN = FKN , for any N = 2 . By getting back now to our classification work, the possible abelian groups at N = 4, that we can use, are the cyclic group Z4, which produces the Fourier matrix F4, and the Klein group K4 = Z2 × Z2, which produces the Walsh matrix W4. The point, however, is that, besides F4,W4, there are many other complex Hadamard matrices at N = 4. Indeed, we can use here the following version of the tensor product construction, coming from Dit¸˘a’spaper [47], involving parameters: Proposition 2.10. If H ∈ MM (T) and K ∈ MN (T) are Hadamard, then so are the following two matrices, for any choice of a parameter matrix Q ∈ MM×N (T): (1) H ⊗Q K ∈ MMN (T), given by (H ⊗Q K)ia,jb = QibHijKab. (2) H Q ⊗ K ∈ MMN (T), given by (H Q ⊗ K)ia,jb = QjaHijKab. These are called right and left Dit¸˘adeformations of H ⊗ K, with parameter Q. HADAMARD MATRICES 25 Proof. The rows Ria of the matrix H ⊗Q K from (1) are indeed pairwise orthogonal: X ¯ ¯ ¯ < Ria,Rkc > = QibHijKab · QkbHkjKcb jb X ¯ = Mδik KabKcb b = MNδikδac As for the rows Lia of the matrix HQ ⊗ K from (2), these are orthogonal as well: X ¯ ¯ ¯ < Lia,Lkc > = QjaHijKab · QjcHkjKcb jb X ¯ = Nδac HijHkj j = MNδikδac Thus, both the matrices in the statement are Hadamard, as claimed. As a first observation, when the parameter matrix is the all-one matrix I ∈ MM×N (T), we obtain in this way the usual tensor product of our matrices: H ⊗I K = HI ⊗ K = H ⊗ K As a non-trivial example now, the right deformations of the Walsh matrix W4 = F2 ⊗F2, p q with arbitrary parameter matrix Q = (r s), are given by: p q p q 1 1 1 1 p −q p −q F2 ⊗Q F2 = ⊗ = 1 −1 p q 1 −1 r s −r −s r s r −s −r s This follows indeed by carefully working out what happens, by using the lexicographic order on the double indices, as explained after Proposition 1.6 above. To be more precise, the usual tensor product W4 = F2 ⊗ F2 appears as follows: ia\jb 00 01 10 11 00 1 1 1 1 W4 = 01 1 −1 1 −1 10 1 1 −1 −1 11 1 −1 −1 1 26 TEO BANICA The corresponding values of the parameters Qib to be inserted are as follows: ia\jb 00 01 10 11 00 Q00 Q01 Q00 Q01 (Qib) = 01 Q00 Q01 Q00 Q01 10 Q10 Q11 Q10 Q11 11 Q10 Q11 Q10 Q11 p q With the notation Q = (r s), this latter matrix becomes: ia\jb 00 01 10 11 00 p q p q (Qib) = 01 p q p q 10 r s r s 11 r s r s Now by pointwise multiplying this latter matrix with the matrix W4 given above, we obtain the announced formula for the deformed tensor product F2 ⊗Q F2. As for the left deformations of W4 = F2 ⊗ F2, once again with arbitrary parameter p q matrix Q = (r s), these are given by a similar formula, as follows: p p r r 1 1 1 1 q −q s −s F2Q ⊗ F2 = ⊗ = 1 −1 p q 1 −1 p p −r −r r s q −q −s s Observe that this latter matrix is transpose to F2 ⊗Q F2. However, this is something accidental, coming from the fact that F2, and so W4 as well, are self-transpose. With the above constructions in hand, we have the following result: Theorem 2.11. The only complex Hadamard matrices at N = 4 are, up to the standard equivalence relation, the matrices 1 1 1 1 s 1 −1 1 −1 F = 4 1 s −1 −s 1 −s −1 s with s ∈ T, which appear as right Dit¸˘adeformations of W4 = F2 ⊗ F2. s Proof. First of all, the matrix F4 is indeed Hadamard, appearing from the construction in Proposition 2.10, assuming that the parameter matrix there Q ∈ M2(T) is dephased: 1 1 Q = 1 s HADAMARD MATRICES 27 Observe also that, conversely, any right Dit¸˘adeformation of W4 = F2 ⊗ F2 is of this p q form. Indeed, if we consider such a deformation, with general parameter matrix Q = (r s) s0 0 as above, by dephasing we obtain an equivalence with F4 , where s = ps/qr: p q p q 1 1 1 1 p −q p −q 1 −1 1 −1 → r s −r −s r/p s/q −r/p −s/q r −s −r s r/p −s/q −r/p s/q 1 1 1 1 1 −1 1 −1 → 1 ps/qr −1 −ps/qr 1 −ps/qr −1 ps/qr s It remains to prove that the matrices F4 are non-equivalent, and that any complex s Hadamard matrix H ∈ M4(T) is equivalent to one of these matrices F4 . But this follows by using the same kind of arguments as in the proof of Proposition 1.7, and from the proof of Proposition 2.6. Indeed, let us first dephase our matrix: 1 1 1 1 1 a b c H = 1 d e f 1 g h i We use now the fact, coming from plane geometry, that the solutions x, y, z, t ∈ T of the equation x + y + z + t = 0 are given by {x, y, z, t} = {p, q, −p, −q}, with p, q ∈ T. In our case, we have 1 + a + d + g = 0, and so up to a permutation of the last 3 rows, our matrix must look at follows, for a certain s ∈ T: 1 1 1 1 1 −1 b c H = 1 s e f 1 −s h i In the case s = ±1 we can permute the middle two columns, then repeat the same reasoning, and we end up with the matrix in the statement. In the case s 6= ±1 we have 1 + s + e + f = 0, and so −1 ∈ {e, f}. Up to a permutation of the last columns, we can assume e = −1, and our matrix becomes: 1 1 1 1 1 −1 b c H = 1 s −1 −s 1 −s h i 28 TEO BANICA Similarly, from 1 − s + h + i = 0 we deduce that −1 ∈ {h, i}. In the case h = −1 our matrix must look as follows, and we are led to the matrix in the statement: 1 1 1 1 1 −1 b c H = 1 s −1 −s 1 −s −1 i As for the remaining case i = −1, here our matrix must look as follows: 1 1 1 1 1 −1 b c H = 1 s −1 −s 1 −s h −1 We obtain from the last column c = s, then from the second row b = −s, then from the third column h = s, and so our matrix must be as follows: 1 1 1 1 1 −1 −s s H = 1 s −1 −s 1 −s s −1 But, in order for the second and third row to be orthogonal, we must have s ∈ R, and so s = ±1, which contradicts our above assumption s 6= ±1. Thus, we are done with the proof of the main assertion. As for the fact that the matrices in the statement are indeed not equivalent, this is standard as well. See [88]. At N = 5 now, the situation is considerably more complicated, with F5 being the only known example, but with the proof of its uniqueness being highly nontrivial. The key technical result here, due to Haagerup [51], is as follows: Proposition 2.12. Given an Hadamard matrix H ∈ M5(T), chosen dephased, 1 1 1 1 1 1 a x ∗ ∗ H = 1 y b ∗ ∗ 1 ∗ ∗ ∗ ∗ 1 ∗ ∗ ∗ ∗ the numbers a, b, x, y must satisfy (x − y)(x − ab)(y − ab) = 0. Proof. This is something quite surprising, and tricky, the proof in [51] being as follows. Let us look at the upper 3-row truncation of H, which is of the following form: 1 1 1 1 1 0 H = 1 a x p q 1 y b r s HADAMARD MATRICES 29 By using the orthogonality of the rows, we have: (1 + a + x)(1 + ¯b +y ¯)(1 +ay ¯ + bx¯) = −(p + q)(r + s)(¯pr +qs ¯ ) On the other hand, by using p, q, r, s ∈ T, we have: (p + q)(r + s)(¯pr +qs ¯ ) = (r + pqs¯ +pqr ¯ + s)(¯r +s ¯) = 1 + pq¯rs¯ +pq ¯ +rs ¯ + rs¯ + pq¯ +pqr ¯ s¯ + 1 = 2Re(1 + pq¯ + rs¯ + pqr¯ s¯) = 2Re[(1 + pq¯)(1 + rs¯)] We conclude that we have the following formula, involving a, b, x, y only: (1 + a + x)(1 + ¯b +y ¯)(1 +ay ¯ + bx¯) ∈ R Now this is a product of type (1+α)(1+β)(1+γ), with the first summand being 1, and with the last summand, namely αβγ, being real as well, as shown by the above general p, q, r, s ∈ T computation. Thus, when expanding, and we are left with: (a + x) + (¯b +y ¯) + (¯ay + bx¯) + (a + x)(¯b +y ¯) + (a + x)(¯ay + bx¯) + (¯b +y ¯)(¯ay + bx¯) ∈ R By expanding all the products, our formula looks as follows: a + x + ¯b +y ¯ +ay ¯ + bx¯ + a¯b + ay¯ + ¯bx + xy¯ + 1 + abx¯ +axy ¯ + b +a ¯¯by +x ¯ +a ¯ + bx¯y¯ ∈ R By removing from this all terms of type z +z ¯, we are left with: a¯b + xy¯ + abx¯ +a ¯¯by +axy ¯ + bx¯y¯ ∈ R Now by getting back to our Hadamard matrix, all this remains true when transposing it, which amounts in interchanging x ↔ y. Thus, we have as well: a¯b +xy ¯ + aby¯ +a ¯¯bx +axy ¯ + bx¯y¯ ∈ R By substracting now the two equations that we have, we obtain: xy¯ − xy¯ + ab(¯x − y¯) +a ¯¯b(y − x) ∈ R Now observe that this number, say Z, is purely imaginary, because Z¯ = −Z. Thus our equation reads Z = 0. On the other hand, we have the following formula: abxyZ = abx2 − aby2 + a2b2(y − x) + xy(y − x) = (y − x)(a2b2 + xy − ab(x + y)) = (y − x)(ab − x)(ab − y) Thus, our equation Z = 0 corresponds to the formula in the statement. 30 TEO BANICA By using the above result, we are led to the following theorem, also from [51]: Theorem 2.13. The only Hadamard matrix at N = 5 is the Fourier matrix, 1 1 1 1 1 1 w w2 w3 w4 2 4 3 F5 = 1 w w w w 3 4 2 1 w w w w 1 w4 w3 w2 w with w = e2πi/5, up to the standard equivalence relation for such matrices. Proof. Assume that have an Hadamard matrix H ∈ M5(T), chosen dephased, and written as in Proposition 2.12, with emphasis on the upper left 2 × 2 subcorner: 1 1 1 1 1 1 a x ∗ ∗ H = 1 y b ∗ ∗ 1 ∗ ∗ ∗ ∗ 1 ∗ ∗ ∗ ∗ We know from Proposition 2.12, applied to H itself, and to its transpose Ht as well, that the entries a, b, x, y must satisfy the following equations: (a − b)(a − xy)(b − xy) = 0 (x − y)(x − ab)(y − ab) = 0 This is of course something very strong, and these equations are actually valid all across the matrix, by permuting rows and columns. The idea will be that by doing some combinatorics, sometimes combined with a few tricks, this will lead to the result. Our first claim is that, by doing some combinatorics, we can actually obtain from this a = b and x = y, up to the equivalence relation for the Hadamard matrices: 1 1 1 1 1 1 a x ∗ ∗ H ∼ 1 x a ∗ ∗ 1 ∗ ∗ ∗ ∗ 1 ∗ ∗ ∗ ∗ Indeed, the above two equations lead to 9 possible cases, the first of which is, as desired, a = b and x = y. As for the remaining 8 cases, here once again things are determined by 2 parameters, and in practice, we can always permute the first 3 rows and 3 columns, and then dephase our matrix, as for our matrix to take the above special form. With this result in hand, the combinatorics of the scalar products between the first 3 rows, and between the first 3 columns as well, becomes something which is quite simple HADAMARD MATRICES 31 to investigate. By doing a routine study here, and then completing it with a study of the lower right 2 × 2 corner as well, we are led to 2 possible cases, as follows: 1 1 1 1 1 1 1 1 1 1 1 a b c d 1 a b c d H ∼ 1 b a d c : H ∼ 1 b a d c 1 c d a b 1 c d b a 1 d c b a 1 d c a b Our claim now is that the first case is in fact not possible. Indeed, we must have: a + b + c + d = −1 2Re(a¯b) + 2Re(cd¯) = −1 2Re(ac¯) + 2Re(bd¯) = −1 2Re(ad¯) + 2Re(bc¯) = −1 Since we have |Re(x)| ≤ 1 for any x ∈ T, we deduce from the second equation that Re(a¯b) ≤ 1/2, and so that the arc length between a, b satisfies θ(a, b) ≥ π/3. The same argument applies to c, d, and to the other pairs of numbers in the last 2 equations. Now since our equations are invariant under permutations of a, b, c, d, we can assume that a, b, c, d are ordered on the circle, and by the above, separated by ≥ π/3 arc lengths. But this implies θ(a, c) ≥ 2π/3 and θ(b, d) ≥ 2π/3, which gives Re(ac¯) ≤ −1/2 and Re(bd¯) ≤ −1/2, which contradicts the third equation. Thus, our claim is proved. Summarizing, we have proved so far that our matrix must be as follows: 1 1 1 1 1 1 a b c d H ∼ 1 b a d c 1 c d b a 1 d c a b We are now in position of finishing. The orthogonality equations are as follows: a + b + c + d = −1 2Re(a¯b) + 2Re(cd¯) = −1 ac¯ + c¯b + bd¯+ da¯ = −1 The third equation can be written in the following equivalent form: Re[(a + b)(¯c + d¯)] = −1 Im[(a − b)(¯c − d¯)] = 0 a+b c+d From a, b, c, d ∈ T we obtain a−b , c−d ∈ iR, so we can find s, t ∈ R such that: a + b = is(a − b) , c + d = it(c − d) 32 TEO BANICA By plugging in these values, our system of equations simplifies, as follows: (a + b) + (c + d) = −1 |a + b|2 + |c + d|2 = 3 (a + b)(¯c + d¯) = −1 Now observe that the last equation implies in particular that we have: |a + b|2 · |c + d|2 = 1 Thus |a + b|2, |c + d|2 must be roots of X2 − 3X + 1 = 0, which gives: (√ √ ) n o 5 + 1 5 − 1 |a + b| , |c + d| = , 2 2 This is very good news, because we are now into 5-th roots of unity. To be more precise, we have 2 cases to be considered, the first one being as follows, with z ∈ T: √ √ 5 + 1 5 − 1 a + b = z , c + d = − z 2 2 From a + b + c + d = −1 we obtain z = −1, and by using this we obtain b =a, ¯ d =c ¯, and then Re(a) = cos(2π/5), Re(c) = cos(π/5), and so we have H ∼ F5. The second case, with a, b and c, d interchanged, this leads to H ∼ F5 as well. The above result is of course something quite impressive. However, at the level of practical conclusions, we can only say that the N = 5 case is something very simple. At N = 6 now, the situation becomes considerably complicated, with lots of “exotic” solutions, and with the structure of the Hadamard manifold X6 being not understood yet. In fact, this manifold X6 looks as complicated as real algebraic manifolds can get. The simplest examples of Hadamard matrices at N = 6 are as follows: Theorem 2.14. We have the following basic Hadamard matrices, at N = 6: (1) The Fourier matrix F6. (2) The Dit¸˘adeformations of F2 ⊗ F3 and of F3 ⊗ F2. q (3) The Haagerup matrix H6 . (4) The Tao matrix T6. Proof. All this is elementary, the idea, and formulae of the matrices, being as follows: (1) This is something that we know well. (2) Consider indeed the dephased Dit¸˘adeformations of F2 ⊗ F3 and F3 ⊗ F2: (rs) (r) F = F ⊗ F ,F s = F ⊗ F 6 2 1 1 1 3 6 3 1 1 2 1 r s 1 r 1 s HADAMARD MATRICES 33 Here r, s are two parameters on the unit circle, r, s ∈ T. In matrix form: 1 1 1 1 1 1 2 2 1 w w 1 w w 2 2 1 w w 1 w w (rs) F6 = 1 r s −1 −r −s 1 wr w2s −1 −wr −w2s 1 w2r ws −1 −w2r −ws As for the other deformation, this is given by: 1 1 1 1 1 1 1 −1 1 −1 1 −1 2 2 r 1 r w wr w w r (s) F6 = 2 2 1 −r w −wr w −w r 1 s w2 w2s w ws 1 −s w2 −w2s w −ws (3) The matrix here, from [51], is as follows, with q ∈ T: 1 1 1 1 1 1 1 −1 i i −i −i q 1 i −1 −i q −q H6 = 1 i −i −1 −q q 1 −i q¯ −q¯ i −1 1 −i −q¯ q¯ −1 i (4) The matrix here, from [90], is as follows, with w = e2πi/3: 1 1 1 1 1 1 2 2 1 1 w w w w 2 2 1 w 1 w w w T6 = 2 2 1 w w 1 w w 1 w2 w2 w 1 w 1 w2 w w2 w 1 q Observe that both H6 and T6 are indeed complex Hadamard matrices. The point with the matrices in Theorem 2.14 is that they are “regular”, in the sense that the scalar products between rows appear in the simplest possible way, namely from vanishing sums of roots of unity, possibly rotated by a scalar. We will be back to this in section 3 below, with a proof that these matrices are the only regular ones, at N = 6. 34 TEO BANICA In the non-regular case now, there are many known constructions at N = 6. Here is one such construction, mildly “exotic”, found by Bj¨orck and Fr¨oberg in [31]: Proposition 2.15. The following is a complex Hadamard matrix, 1 ia −a −i −a¯ ia¯ ia¯ 1 ia −a −i −a¯ −a¯ ia¯ 1 ia −a −i BF6 = −i −a¯ ia¯ 1 ia −a −a −i −a¯ ia¯ 1 ia ia −a −i −a¯ ia¯ 1 √ where a ∈ T is one of the roots of a2 + ( 3 − 1)a + 1 = 0. Proof. Observe that the matrix in the statement is circulant, in the sense the rows appear by cyclically permuting the first row. Thus, we only have to√ check that the first row is 2 orthogonal to the other 5 rows. But this follows from a + ( 3 − 1)a + 1 = 0. The obvious question here is perhaps on how Bj¨orck and Fr¨oberg were able to construct the above matrix (!) This was done via some general theory for the circulant Hadamard matrices, and some computer simulations. We will discuss this in section 6 below. Further study in the N = 6 case leads to a number of horrors, of real algebraic geometric flavor, and we have here, as an illustrating example, the following result from [25]: Theorem 2.16. The self-adjoint 6 × 6 Hadamard matrices are, up to equivalence 1 1 1 1 1 1 1 −1x ¯ −y −x¯ y q 1 x −1 t −t −x BN6 = ¯ ¯ 1 −y¯ t −1y ¯ −t 1 −x −t¯ y 1z ¯ 1y ¯ −x¯ −t z 1 with x, y, z, t ∈ T depending on a parameter q ∈ T, in a very complicated way. Proof. The study here can be done via a lot of work, and tricks, in the spirit of the Haagerup classification result at N = 5, and the equations are as follows: √ 1 + 2q + q2 − 2p1 + 2q + 2q3 + q4 x = 1 + 2q − q2 y = q 1 + 2q − q2 z = q(−1 + 2q + q2) √ 1 + 2q + q2 − 2p1 + 2q + 2q3 + q4 t = −1 + 2q + q2 All this is very technical, and we refer here to [25]. HADAMARD MATRICES 35 There are many other examples at N = 6, and no classification known. See [68]. Let us discuss now the case N = 7. We will restrict the attention to case where the combinatorics comes from roots of unity. We use the following result, from [87]: Theorem 2.17. If H ∈ MN (±1) with N ≥ 8 is dephased symmetric Hadamard, and √ (1 ± i N − 5)2 w = N − 4 then the following procedure yields a complex Hadamard matrix M ∈ MN−1(T): (1) Erase the first row and column of H. (2) Replace all diagonal 1 entries with −w. (3) Replace all off-diagonal −1 entries with w. Proof. We know from the proof of Proposition 1.8 that the scalar product between any two rows of H, normalized as there, appears as follows: N N N N P = · 1 · 1 + · 1 · (−1) + · (−1) · 1 + · (−1) · (−1) = 0 4 4 4 4 Let us peform now the above operations (1,2,3), in reverse order. When replacing −1 → w, all across the matrix, the above scalar product becomes: N N N N N P 0 = · 1 · 1 + · 1 · w¯ + · w · 1 + · (−1) · (−1) = (1 + Re(w)) 4 4 4 4 2 Now when adjusting the diagonal via w → −1 back, and 1 → −w, this amounts in adding the quantity −2(1 + Re(w)) to our product. Thus, our product becomes: N N − 4 6 − N P 00 = − 2 (1 + Re(w)) = 1 + = 1 2 2 N − 4 Finally, erasing the first row and column amounts in substracting 1 from our scalar 000 product. Thus, our scalar product becomes P = 1 − 1 = 0, and we are done. Observe that the number w in the above statement is a root of unity precisely at N = 8, where the only matrix satisfying the conditions in the statement is the Walsh matrix W8. So, let us apply, as in [87], the above construction to this matrix: 1 1 1 1 1 1 1 1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 1 −1 1 −1 1 −1 1 −1 ∗ −1 1 w 1 w 1 w 1 1 −1 −1 1 1 −1 −1 ∗ 1 −1 w 1 1 w w 1 −1 −1 1 1 −1 −1 1 ∗ w w −w 1 w w 1 → 1 1 1 1 −1 −1 −1 −1 ∗ 1 1 1 −1 w w w 1 −1 1 −1 −1 1 −1 1 ∗ w 1 w w −w w 1 1 −1 −1 −1 −1 −1 1 1 ∗ 1 w w w w −w 1 1 −1 −1 1 −1 1 1 −1 ∗ w w 1 w 1 1 −1 The matrix on the right is the Petrescu matrix P7, found in [77]. Thus, we have: 36 TEO BANICA Theorem 2.18. P7 is the unique matrix formed by roots of unity that can be obtained by the Sz¨oll˝osiconstruction. It appears at N = 8, from H = W8. Its formula is −w if (ijk) = (abc), ia + jb + kc = 0(2) (P7)ijk,abc = w if (ijk) 6= (abc), ia + jb + kc 6= 0(2) (−1)ia+jb+kc otherwise where w = e2πi/3, and with the indices belonging to the set {0, 1}3 − {(0, 0, 0)}. ij Proof. We know that the Sz¨oll˝osiconstruction maps W8 → P7. Since (F2)ij = (−1) , we ia+jb+kc have (W8)ijk,abc = (−1) , and this gives the formula in the statement. Now observe that we are in the quite special situation H = F2 ⊗ K, with K being dephased and symmetric. Thus, we can search for a one-parameter affine deformation K(q) which is dephased and symmetric, and then build the following matrix: K(q) K H(q) = K −K(¯q) In our case, such a deformation K(q) = W4(q) can be obtained by putting the q parameters in the 2 × 2 middle block. Now by performing the Sz¨oll˝osiconstruction, with the parameters q, q¯ left untouched, we obtain the parametric Petrescu matrix [77]: Theorem 2.19. The following is a complex Hadamard matrix, −q q w 1 w 1 w q −q w 1 1 w w w w −w 1 w w 1 q 1 1 1 −1 w w w P7 = w 1 w w −qw¯ qw¯ 1 1 w w w qw¯ −qw¯ 1 w w 1 w 1 1 −1 where w = e2πi/3, and q ∈ T. Proof. This follows from the above considerations, or from a direct verification of the 2 orthogonality of the rows, which uses either 1 − 1 = 0, or 1 + w + w = 0. q Observe that the above matrix P7 has the property of being “regular”, in the sense that the scalar products between rows appear from vanishing sums of roots of unity, possibly rotated by a scalar. We will be back to this in the next section, with the conjectural q statement that F7,P7 are the only regular Hadamard matrices at N = 7. HADAMARD MATRICES 37 3. Roots of unity Many interesting examples of complex Hadamard matrices H ∈ MN (T), including the real ones H ∈ MN (±1), have as entries roots of unity, of finite order. We discuss here this case, and more generally the “regular” case, where the combinatorics of the scalar products between the rows comes from vanishing sums of roots of unity. Let us begin with the following definition, going back to the work in [34]: Definition 3.1. An Hadamard matrix is called of Butson type if its entries are roots of unity of finite order. The Butson class HN (l) consists of the Hadamard matrices H ∈ MN (Zl) where Zl is the group of the l-th roots of unity. The level of a Butson matrix H ∈ MN (T) is the smallest integer l ∈ N such that H ∈ HN (l). As basic examples, we have the real Hadamard matrices, which form by definition the Butson class HN (2). The Fourier matrices are Butson matrices as well, because we have FN ∈ HN (N), and more generally FG ∈ HN (l), with N = |G|, and with l ∈ N being the smallest common order of the elements of G. There are many other examples of such matrices, as for instance those in Theorem 2.14, at 1 values of the parameters. Generally speaking, the main question regarding the Butson matrices is that of under- standing when HN (l) 6= 0, via a theorem providing obstructions, and then a conjecture stating that these obstructions are the only ones. Let us begin with: Proposition 3.2 (Sylvester obstruction). The following holds, HN (2) 6= ∅ =⇒ N ∈ {2} ∪ 4N due to the orthogonality of the first 3 rows. Proof. This is something that we know from section 1, with the obstruction, going back to Sylvester’s paper [85], being explained in Proposition 1.8 above. The above obstruction is fully satisfactory, because according to the Hadamard Con- jecture, its converse should hold. Thus, we are fully done with the case l = 2. Our purpose now will be that of finding analogous statements at l ≥ 3, theorem plus conjecture. At very small values of l this is certainly possible, and in what regards the needed obstructions, we can get away with the following simple fact, from [34], [98]: Proposition 3.3. For a prime power l = pa, the vanishing sums of l-th roots of unity λ1 + ... + λN = 0 , λi ∈ Zl appear as formal sums of rotated full sums of p-th roots of unity. 38 TEO BANICA Proof. Consider indeed the full sum of p-th roots of unity, taken in a formal sense: p X S = (e2πi/p)k k=1 2πi/l r r Let also w = e , and for r ∈ {1, 2, . . . , l/p} denote by Sp = w · S the above sum, rotated by wr. We must show that any vanishing sum of l-th roots of unity appears as a r sum of such quantities Sp, with all this taken of course in a formal sense. For this purpose, consider the following map, which assigns to the abstract elements of the group ring Z[Zl] their precise numeric values, inside Z(w) ⊂ C: Φ: Z[Zl] → Z(w) r Our claim is that the elements {Sp} form a basis of ker Φ. Indeed, we obviously have r r Sp ∈ ker Φ. Also, these elements are linearly independent, because the support of Sp a−1 contains a unique element of the subset {1, 2, . . . , p } ⊂ Zl, namely the element r ∈ Zl, r so all the coefficients of a vanishing linear combination of sums Sp must vanish. r Thus, we are left with proving that ker Φ is spanned by {Sp}. For this purpose, let us recall that the minimal polynomial of w is as follows: pa X − 1 a−1 a−1 a−1 = 1 + Xp + X2p + ... + X(p−1)p Xpa−1 − 1 But this shows that ker Φ has dimension pa − (pa − pa−1) = pa−1, and since this is r exactly the number of the sums Sp, this finishes the proof of our claim. P r Thus, any vanishing sum of l-th roots of unity must be of the form ±Sp, and the above support considerations show the coefficients must be positive, as desired. We can now formulate a result in the spirit of Proposition 3.2, as follows: Proposition 3.4 (Butson obstruction). The following holds, a HN (p ) 6= ∅ =⇒ N ∈ pN due to the orthogonality of the first 2 rows. Proof. This follows indeed from Proposition 3.3, because the scalar product between the first 2 rows of our matrix is a vanishing sum of l-th roots of unity. WIth these obstructions in hand, we can discuss the case l ≤ 5, as follows: Theorem 3.5. We have the following results, (1) HN (2) 6= ∅ =⇒ N ∈ {2} ∪ 4N, (2) HN (3) 6= ∅ =⇒ N ∈ 3N, (3) HN (4) 6= ∅ =⇒ N ∈ 2N, (4) HN (5) 6= ∅ =⇒ N ∈ 5N, with in cases (1,3), a solid conjecture stating that the converse should hold as well. HADAMARD MATRICES 39 Proof. In this statement (1) is the Sylvester obstruction, and (2,3,4) are particular cases of the Butson obstruction. As for the last assertion, which is of course something rather informal, but which is important for our purposes, the situation is as follows: (1) Here, as already mentioned, we have the Hadamard Conjecture, which comes with very solid evidence, as explained in section 1 above. (2) Here we have an old conjecture, dealing with complex Hadamard matrices over {±1, ±i}, going back to the work in [93], and called Turyn Conjecture. At l = 3 the situation is quite complicated, due to the following result, from [40]: Proposition 3.6 (de Launey obstruction). The following holds, 2πi/l 2 N HN (l) 6= ∅ =⇒ ∃ d ∈ Z[e ], |d| = N due to the orthogonality of all N rows. In particular, we have 5|N =⇒ HN (6) = ∅ and so H15(3) = ∅, which shows that the Butson obstruction is too weak at l = 3. Proof. The obstruction follows from the unitarity condition HH∗ = N for the complex Hadamard matrices, by applying the determinant to it, which gives: |det(H)|2 = N N Regarding the second assertion, let w = e2πi/3, and assume that d = a + bw + cw2 with a, b, c ∈ Z satisfies |d|2 = 0(5). We have the following computation: |d|2 = (a + bw + cw2)(a + bw2 + cw) = a2 + b2 + c2 − ab − bc − ac 1 = [(a − b)2 + (b − c)2 + (c − a)2] 2 Thus our condition |d|2 = 0(5) leads to the following system, modulo 5: x + y + z = 0 , x2 + y2 + z2 = 0 But this system has no solutions. Indeed, let us look at x2 +y2 +z2 = 0. If this equality appears as 0 + 0 + 0 = 0 we can divide x, y, z by 5 and redo the computation, and if not, this equality can only appear as 0 + 1 + (−1) = 0. Thus, modulo permutations, we must have x = 0, y = ±1, z = ±2, which contradicts x + y + z = 0. Finally, the last assertion follows from H15(3) ⊂ H15(6) = ∅. At l = 5 now, things are a bit unclear, with the converse of Theorem 3.5 (4) being something viable, at the conjectural level, at least to our knowledge. At l = 6 the situation becomes again complicated, as follows: 40 TEO BANICA Proposition 3.7 (Haagerup obstruction). The following holds, due to Haagerup’s N = 5 classification result, involving the orthogonality of all 5 rows of the matrix: H5(l) 6= ∅ =⇒ 5|l In particular we have H5(6) = ∅, which follows by the way from the de Launey obstruction as well, in contrast with the fact that we generally have HN (6) 6= ∅. Proof. In this statement the obstruction H5(l) = ∅ =⇒ 5|l comes indeed from Haagerup’s classification result, explained in Theorem 2.13 above. As for the last assertion, this is something very informal, the situation at small values of N being as follows: – At N = 2, 3, 4 we have the matrices F2,F3,W4. 1 – At N = 6, 7, 8, 9 we have the matrices F6,P7 ,W8,F3 ⊗ F3. – At N = 10 we have the following matrix, found in [13] by using a computer, and written in logarithmic form, with k standing for e2kπi/6: 0 0 0 0 0 0 0 0 0 0 0 4 1 5 3 1 3 3 5 1 0 1 2 3 5 5 1 3 5 3 0 5 3 2 1 5 3 5 3 1 6 0 3 5 1 4 1 1 5 3 3 X10 = 0 3 3 3 3 3 0 0 0 0 0 1 1 5 3 4 3 0 2 4 0 1 5 3 5 2 4 3 2 0 0 5 3 5 1 2 0 2 3 4 0 3 5 1 1 4 4 2 0 3 We refer to [13] for more details on this topic. All this is not good news. Indeed, there is no hope of conjecturally solving our HN (l) 6= ∅ problem in general, because this would have to take into account, and in a simple and conceptual way, both the subtle arithmetic consequences of the de Launey obstruction, and the Haagerup classification result at N = 5, and this is something not feasible. In order to further comment on these difficulties, let us discuss now a generalization of Proposition 3.3 above, and of the related Butson obstruction from Proposition 3.4, which has been our main source of obstructions, so far. Let us start with: Definition 3.8. A cycle is a full sum of roots of unity, possibly rotated by a scalar, l X k 2πi/l C = q w , w = e , q ∈ T k=1 and taken in a formal sense. A sum of cycles is a formal sum of cycles. HADAMARD MATRICES 41 The actual sum of a cycle, or of a sum of cycles, is of course 0. This is why the word “formal” is there, for reminding us that we are working with formal sums. As an example, here is a sum of cycles, with w = e2πi/6, and with |q| = 1: 1 + w2 + w4 + qw + qw4 = 0 We know from Proposition 3.3 above that any vanishing sum of l-th roots of unity must be a sum of cycles, at least when l = pa is a prime power. However, this is not the case in general, the simplest counterexample being as follows, with w = e2πi/30: w5 + w6 + w12 + w18 + w24 + w25 = 0 The following deep result on the subject is due to Lam and Leung [66]: a1 ak Theorem 3.9. Let l = p1 . . . pk , and assume that λi ∈ Zl satisfy λ1 + ... + λN = 0. P (1) λi is a sum of cycles, with Z coefficients. P (2) If k ≤ 2 then λi is a sum of cycles (with N coefficients). P (3) If k ≥ 3 then λi might not decompose as a sum of cycles. P (4) λi has the same length as a sum of cycles: N ∈ p1N + ... + pkN. Proof. This is something that we will not really need in what follows, but that we included here, in view of its importance. The idea of the proof is as follows: (1) This is a well-known result, which follows from basic number theory, by using arguments in the spirit of those in the proof of Proposition 3.3 above. (2) This is something that we already know at k = 1, from Proposition 3.3. At k = 2 the proof is more technical, along the same lines. See [66]. (3) The smallest possible l potentially producing a counterexample is l = 2 · 3 · 5 = 30, and we have here indeed the sum given above, with w = e2πi/30. (4) This is a deep result, due to Lam and Leung, relying on advanced number theory knowledge. We refer to their paper [66] for the proof. As a consequence of the above result, we have the following generalization of the Butson obstruction, which is something final and optimal on this subject: a1 ak Theorem 3.10 (Lam-Leung obstruction). Assuming l = p1 . . . pk , the following must hold, due to the orthogonality of the first 2 rows: HN (l) 6= ∅ =⇒ N ∈ p1N + ... + pkN In the case k ≥ 2, the latter condition is automatically satisfied at N >> 0. Proof. Here the first assertion, which generalizes the l = pa obstruction from Proposition 3.4 above, comes from Theorem 3.9 (4), applied to the vanishing sum of l-th roots of unity coming from the scalar product between the first 2 rows. As for the second assertion, this is something well-known, coming from basic number theory. 42 TEO BANICA Summarizing, our study so far of the condition HN (l) 6= ∅ has led us into an optimal obstruction coming from the first 2 rows, namely the Lam-Leung one, then an obstruction coming from the first 3 rows, namely the Sylvester one, and then two subtle obstructions coming from all N rows, namely the de Launey one, and the Haagerup one. As an overall conclusion, by contemplating all these obstructions, nothing good in relation with our problem HN (l) 6= ∅ is going on at small N. So, as a natural and more modest objective, we should perhaps try instead to solve this problem at N >> 0. The point indeed is that everything simplifies at N >> 0, with some of the above obstructions dissapearing, and with some other known obstructions, not to be discussed here, dissapearing as well. We are therefore led to the following statement: Conjecture 3.11 (Asymptotic Butson Conjecture (ABC)). The following equivalences should hold, in an asymptotic sense, at N >> 0, (1) HN (2) 6= ∅ ⇐⇒ 4|N, a a (2) HN (p ) 6= ∅ ⇐⇒ p|N, for p ≥ 3 prime power, (3) HN (l) 6= ∅ ⇐⇒ ∅, for l ∈ N not a prime power, modulo the de Launey obstruction, |d|2 = N N for some d ∈ Z[e2πi/l]. In short, our belief is that when imposing the condition N >> 0, only the Sylvester, Butson and de Launey obstructions survive. This is of course something quite nice, but in what regards a possible proof, there is probably no way. Indeed, our above conjecture generalizes the HC in the N >> 0 regime, which is something beyond reach. One interesting idea, however, in dealing with such questions, coming from the de Launey-Levin result from [44], explained in section 1, is that of looking at the partial Butson matrices, at N >> 0. Observe in particular that restricting the attention to the rectangular case, and this not even in the N >> 0 regime, would make dissapear the de Launey obstruction from the ABC, which uses the orthogonality of all N rows. We will discuss this later on, at the end of this section. For a number of related considerations, we refer as well to the papers [40], [43]. Getting away now from all this arithmetic madness, let us discuss now, as a more concrete thing, the classification of the regular complex Hadamard matrices of small order. The definition here, which already appeared in the above, is as follows: Definition 3.12. A complex Hadamard matrix H ∈ MN (T) is called regular if the scalar products between rows decompose as sums of cycles. We should mention that there is some notational clash here, with this notion being sometimes used in order to designate the bistochastic matrices. In this book we use the above notion of regularity, and we call bistochastic the bistochastic matrices. Our purpose in what follows will be that of showing that the notion of regularity can lead to full classification results at N ≤ 6, and perhaps at N = 7 too, and all this while covering most of the interesting complex Hadamard matrices that we met, so far. HADAMARD MATRICES 43 As a first observation, supporting this last claim, we have the following result: Proposition 3.13. The following complex Hadamard matrices are regular: s (1) The matrices at N ≤ 5, namely F2,F3,F4 ,F5. r (rs) (s) q (2) The main examples at N = 6, namely F6 ,F6 ,H6 ,T6. q (3) The main examples at N = 7, namely F7,P7 . Proof. The Fourier matrices FN are all regular, with the scalar products between rows appearing as certain sums of full sums of l-th roots of unity, with l|N. As for the other matrices appearing in the statement, with the convention that “cycle structure” means the length of the cycles in the regularity property, the situation is as follows: s (1) F4 has cycle structure 2 + 2, and this because the verification of the Hadamard condition is always based on the formula 1 + (−1) = 0, rotated by scalars. r (rs) (s) (2) F6 ,F6 have mixed cycle structure 2 + 2 + 2/3 + 3, in the sense that both cases q appear, H6 has cycle structure 2 + 2 + 2, and T6 has cycle structure 3 + 3. q 2 (3) P7 has cycle structure 3+2+2, its Hadamard property coming from 1+w+w = 0, 2πi/3 with w = e , and from 1 + (−1) = 0, applied twice, rotated by scalars. Let us discuss now the classification of regular matrices. We first have: Theorem 3.14. The regular Hadamard matrices at N ≤ 5 are s F2,F3,F4 ,F5 up to the equivalence relation for the complex Hadamard matrices. Proof. This is something that we already know, coming from the classification results from section 2, and from Proposition 3.13 (1). However, and here comes our point, proving this result does not need in fact all this, the situation being as follows: (1) At N = 2 the cycle structure can be only 2, and we obtain F2. (2) At N = 3 the cycle structure can be only 3, and we obtain F3. s (3) At N = 4 the cycle structure can be only 2 + 2, and we obtain F4 . (4) At N = 5 some elementary combinatorics shows that the cycle structure 3 + 2 is excluded. Thus we are left with the cycle structure 5, and we obtain F5. Let us discuss now the classification at N = 6. The result here, from [13], states that r (rs) (s) q the above matrices F6 ,F6 ,H6 ,T6 are the only solutions. The proof of this fact is quite long and technical, but we will present here its main ideas. Let us start with: Proposition 3.15. The regular Hadamard matrices at N = 6 fall into 3 classes: (1) Cycle structure 3 + 3, with T6 being an example. q (2) Cycle structure 2 + 2 + 2, with H6 being an example. r (rs) (s) (3) Mixed cycle structure 3 + 3/2 + 2 + 2, with F6 ,F6 being examples. Proof. This is a bit of an empty statement, with the above (1,2,3) possibilities being the only ones, and with the various examples coming from Proposition 3.13 (2). 44 TEO BANICA In order to do the classification, we must prove that the examples in (1,2,3) are the only ones. Let us start with the Tao matrix. The result here is as follows: Proposition 3.16. The matrix T6 is the only one with cycle structure 3 + 3. Proof. The proof of this fact, from [13], is quite long and technical, the idea being that of studying first the 3 × 6 case, then the 4 × 6 case, and finally the 6 × 6 case. So, consider first a partial Hadamard matrix A ∈ M3×6(T), with the scalar products between rows assumed to be all of type 3+3. By doing some elementary combinatorics, one 2 can show that, modulo equivalence, either all the entries of A belong to Z3 = {1, w, w }, or A has the following special form, for certain parameters r, s ∈ T: 1 1 1 1 1 1 A = 1 w w2 r wr w2r 1 w2 w s w2s ws With this in hand, we can now investigate the 4×6 case. Assume indeed that we have a partial Hadamard matrix B ∈ M4×6(T), with the scalar products between rows assumed to be all of type 3 + 3. By looking at the 4 submatrices A(1),A(2),A(3),A(4) obtained from B by deleting one row, and applying the above 3 × 6 result, we are led, after doing some combinatorics, to the conclusion that all the possible parameters dissapear: B ∈ M4×6(Z3) With this result in hand, we can go now for the general case. Indeed, an Hadamard matrix M ∈ M6(T) having cycle structure 3 + 3 must be as follows: M ∈ M6(Z3) But the study here is elementary, with T6 as the only solution. See [13]. Regarding now the Haagerup matrix, the result is similar, as follows: q Proposition 3.17. The matrix H6 is the only one with cycle structure 2 + 2 + 2. Proof. The proof here, from [13], uses the same idea as in the proof of Proposition 3.16. The study of the 3 × 6 partial Hadamard matrices with cycle structure 2 + 2 + 2 leads, up to equivalence, to the following 4 solutions, with q ∈ T being a parameter: 1 1 1 1 1 1 A1 = 1 −i 1 i −1 −1 1 −1 i −i q −q 1 1 1 1 1 1 A2 = 1 1 −1 i −1 −i 1 −1 q −q iq −iq HADAMARD MATRICES 45 1 1 1 1 1 1 A3 = 1 −1 i −i q −q 1 −i i −1 −q q 1 1 1 1 1 1 A4 = 1 −i −1 i q −q 1 −1 −q −iq iq q With this result in hand, we can go directly for the 6 × 6 case. Indeed, a careful examination of the 3×6 submatrices, and of the way that different parameters can overlap vertically, shows that our matrix must have a 3 × 3 block decomposition as follows: ABC M = D xE yF G zH tI Here A, . . . , I are 2 × 2 matrices over {±1, ±i}, and x, y, z, t are in {1, q}. A more careful examination shows that the solution must be of the following form: ABC M = D E qF G qH qI More precisely, the matrix must be as follows: 1 1 1 1 1 1 1 1 −i i −1 −1 1 i −1 −i −q q M = 1 −i i −1 −iq iq 1 −1 q −iq iq −q 1 −1 −q iq q −iq q But this matrix is equivalent to H6 , and we are done. See [13]. Regarding now the mixed case, where both 2 + 2 + 2 and 3 + 3 situations can appear, this is a bit more complicated, and requires some preliminary discussion. We can associate to any mixed Hadamard matrix M ∈ M6(C) its “row graph”, having the 6 rows as vertices, and with each edge being called “binary” or “ternary”, depending on whether the corresponding scalar product is of type 2 + 2 + 2 or 3 + 3. With this convention, we have the following result: Proposition 3.18. The row graph of a mixed matrix M ∈ M6(C) can be: (1) Either the bipartite graph having 3 binary edges. (2) Or the bipartite graph having 2 ternary triangles. 46 TEO BANICA Proof. This is once again something a bit technical, from [13], the idea being as follows. Let X be the row graph in the statement. By doing some combinatorics, quite long but of very elementary type, we are led to the following conclusions about X: – X has no binary triangle. – X has no ternary square. – X has at least one ternary triangle. With these results in hand, we see that there are only two types of squares in our graph X, namely those having 1 binary edge and 5 ternary edges, and those consisting of a ternary triangle, connected to the 4-th point with 3 binary edges. By looking at pentagons, then hexagons that can be built with these squares, we see that the above two types of squares cannot appear at the same time, at that at the level of hexagons, we have the two solutions in the statement. See [13]. We can now complete our classification at N = 6, as follows: r (rs) (s) Proposition 3.19. The matrices F6 ,F6 are the only ones with mixed cycle structure. Proof. According to Proposition 3.18, we have two cases: (1) Assume first that the row graph is the bipartite one with 3 binary edges. By permuting the rows, the upper 4 × 6 submatrix of our matrix must be as follows: 1 1 1 1 1 1 1 w w2 r wr w2r B = 2 2 1 w w s w s ws 1 1 1 t t t Now since the scalar product between the first and the fourth row is binary, we must have t = −1, so the solution is: 1 1 1 1 1 1 1 w w2 r wr w2r B = 2 2 1 w w s w s ws 1 1 1 −1 −1 −1 We can use the same argument for finding the fifth and sixth row, by arranging the matrix formed by the first three rows such as the second, respectively third row consist only of 1’s. This will make appear some parameters of the form w, w2, r, s in the extra (rs) row, and we obtain in this way a matrix which is equivalent to F6 . See [13]. (2) Assume now that the row graph is the bipartite one with 2 ternary triangles. By permuting the rows, the upper 4 × 6 submatrix of our matrix must be as follows: 1 1 1 1 1 1 1 1 w w w2 w2 B = 2 2 1 1 w w w w 1 −1 r −r s −s HADAMARD MATRICES 47 We can use the same argument for finding the fifth and sixth row, and we conclude that the matrix is of the following type: 1 1 1 1 1 1 2 2 1 1 w w w w 2 2 1 1 w w w w M = 1 −1 r −r s −s 1 −1 a −a b −b 1 −1 c −c d −d Now since the last three rows must form a ternary triangle, we conclude that the matrix must be of the following form: 1 1 1 1 1 1 2 2 1 1 w w w w 2 2 1 1 w w w w M = 1 −1 r −r s −s 1 −1 wr −wr w2s −w2s 1 −1 w2r −w2r ws −ws r (s) But this matrix is equivalent to F6 , and we are done. See [13]. Summing up all the above, we have proved the following theorem: Theorem 3.20. The regular complex Hadamard matrices at N = 6 are: r (rs) (s) (1) The deformations F6 ,F6 of the Fourier matrix F6. q (2) The Haagerup matrix H6 . (3) The Tao matrix T6. Proof. This follows indeed from the trichotomy from Proposition 3.15, and from the results in Proposition 3.16, Proposition 3.17 and Proposition 3.19. See [13]. All this is quite nice, and our belief is that the N = 7 classification is doable as well. Here we have 3 possible cycle structures, namely 3 + 2 + 2, 5 + 2, 7, and some elementary number theory shows that 5 + 2 is excluded, and that 3 + 2 + 2 and 7 cannot interact. Thus we have a dichotomy, and our conjecture is as follows: Conjecture 3.21. The regular complex Hadamard matrices at N = 7 are: (1) The Fourier matrix F7. q (2) The Petrescu matrix P7 . Regarding (1), one can show indeed that F7 is the only matrix having cycle structure 7, with this being related to some more general results from [55]. As for (2), the problem q is that of proving that P7 is the only matrix having cycle structure 3 + 2 + 2. The computations here are unfortunately far more involved than those at N = 6, briefly presented above, and finishing the classification work here is not an easy question. 48 TEO BANICA As a conclusion to all this, when imposing the regularity condition, things simplify a bit, with respect to the general case, according to a kind of N → N + 1 rule. To be more precise, the difficulties in the general case are basically of real algebraic geometry nature, and can be labeled as easy at N ≤ 4, hard at N = 5, and not solved yet at N = 6. As for the regular case, here the difficulties are basically of design theory nature, and can be labeled as easy at N ≤ 5, hard at N = 6, and not solved yet at N = 7. Besides the classification questions, there are as well a number of theoretical questions in relation with the notion of regularity, that we believe to be very interesting. We have for instance the following conjecture, going back to [13], and then to [22]: Conjecture 3.22 (Regularity Conjecture). The following hold: (1) Any Butson matrix H ∈ MN (C) is regular. (2) Any regular matrix H ∈ MN (C) is an affine deformation of a Butson matrix. In other words, the first conjecture is that a “tricky vanishing sum” of roots of unity, like the l = 30 one given after Definition 3.8 above, cannot be used in order to construct a complex Hadamard matrix. This is a quite difficult question, coming however with substantial computer evidence. We have no idea on how to approach it. See [13]. As for the second conjecture, this simply comes from the known examples of regular Hadamard matrices, which all appear from certain Butson matrices, by inserting param- eters, in an affine way. This conjecture is from [22], and we will further discuss the notion of affine deformation, with some general results on the subject, in section 4 below. We would like to end this section, which was depressingly algebraic and difficult, with no simple and conceptual result in sight, by doing some analysis. As explained after Conjecture 3.11 above (ABC), one way of getting into analysis, in connection with root of unity questions, is that of looking at the partial Butson matrices, at N >> 0. The idea here comes of course from the de Launey-Levin counting result from [44], explained in section 1 above. Let us first discuss the prime power case. We have: Proposition 3.23. When q = pk is a prime power, the standard form of the dephased partial Butson matrices at M = 2 is 1 1 ... 1 ...... 1 1 ... 1 q/p−1 q−q/p q−q/p+1 q−1 H = 1 w . . . w ...... w w . . . w |{z} |{z} | {z } | {z } | {z } |{z} a1 a2 aq/p a1 a2 aq/p 2πi/q where w = e and where a1, . . . , aq/p ∈ N are multiplicities, summing up to N/p. k Proof. Indeed, it is well-known that for q = p the solutions of λ1 + ... + λN = 0 with λi ∈ Zq are, up to permutations of the terms, exactly those in the statement. Our objective will be to count the matrices in Proposition 3.23. We will need: HADAMARD MATRICES 49 Proposition 3.24. We have the estimate s p s(p−1) X n pn s ' s s−1 (s−1)(p−1) a1, . . . , as p (2πn) a1+...+as=n in the n → ∞ limit. Proof. This is proved by Richmond and Shallit in [81] at p = 2, and the proof in the general case, p ∈ N, is similar. More precisely, let us denote by csp the sum on the left. n √ By setting ai = s + xi n and then by using the various formulae in [81], we obtain: csp s ! pn (1−s)p sp sp X 2 ' s (2πn) 2 s 2 exp − x 2 i i=1 Z n Z n s ! (1−s)p sp sp X pn 2 2 2 ' s (2πn) s ... exp − xi da1 . . . das−1 0 0 2 | {z } i=1 s−1 2 Z n Z n s−1 s−1 ! (1−s)p sp s−1 sp X sp X pn 2 2 2 2 = s (2πn) s n ... exp − xi − xi dx1 . . . dxs−1 0 0 2 2 | {z } i=1 i=1 s−1 1−s pn (1−s)p sp s−1 s−1 − 1 sp 2 = s (2πn) 2 s 2 n 2 × π 2 s 2 2 1−s pn (1−s)p sp − 1 + 1−s p 2 = s (2πn) 2 s 2 2 2 2πn pn (1−s)(p−1) sp−s 1−s = s (2πn) 2 s 2 p 2 Thus we have obtained the formula in the statement, and we are done. Now with Proposition 3.24 in hand, we can prove: Theorem 3.25. When q = pk is a prime power, the probability for a randomly chosen M ∈ M2×N (Zq), with N ∈ pN, N → ∞, to be partial Butson is: s 2− q q− q p p q p P ' 2 q− q (2πN) p r q/2 q pp k q/2 In particular, for q = p prime, P2 ' (2πN)p−1 . Also, for q = 2 , P2 ' 2 2πN . Proof. First, the probability PM for a random M ∈ MM×N (Zq) to be PBM is: 1 P = #PBM M qMN M×N 50 TEO BANICA Thus, according to Proposition 3.23, we have the following formula: 1 X N P2 = N q a1 . . . a1 ...... aq/p . . . aq/p a1+...+aq/p=N/p | {z } | {z } p p p 1 N X N/p = N q N/p . . . N/p a1 . . . aq/p | {z } a1+...+aq/p=N/p p p 1 N 1 X N/p = N × N p N/p . . . N/p (q/p) a1 . . . aq/p | {z } a1+...+aq/p=N/p p Now by using the Stirling formula for the left term, and Proposition 3.24 with s = q/p and n = N/p for the right term, we obtain: s q (p−1) r pp (q/p) p P = × 2 p−1 q −1 ( q −1)(p−1) (2πN) p p (2πN/p) p s p− q (p−1)− q +1+( q −1)(p−1) q (p−1) p p p p q p = p−1+( q −1)(p−1) (2πN) p s 2− q q− q p p q p = q− q (2πN) p Thus we have obtained the formula in the statement, and we are done. k1 k2 Let us discuss now the case where M = 2 and q = p1 p2 has two prime factors. We first examine the simplest such case, namely q = p1p2, with p1, p2 primes: Proposition 3.26. When q = p1p2 is a product of distinct primes, the standard form of the dephased partial Butson matrices at M = 2 is 1 1 ... 1 ...... 1 1 ... 1 p −1 q−p q−p +1 q−1 H = 1 w . . . w 2 ...... w 2 w 2 . . . w |{z} |{z} | {z } | {z } | {z } |{z} A11 A12 A1p2 Ap11 Ap12 Ap1p2 2πi/q where w = e , and A ∈ Mp1×p2 (N) is of the form Aij = Bi + Cj, with Bi,Cj ∈ N. Proof. We use the fact that for q = p1p2 any vanishing sum of q-roots of unity decomposes as a sum of cycles. Now if we denote by Bi,Cj ∈ N the multiplicities of the various p2- cycles and p1-cycles, then we must have Aij = Bi + Cj, as claimed. Regarding the matrices of type Aij = Bi +Cj, when taking them over integers, Bi,Cj ∈ Z, these form a vector space of dimension p1 + p2 − 1. Given A ∈ Mp1×p2 (Z), the “test” for deciding if we have Aij = Bi + Cj or not is Aij + Akl = Ail + Ajk. HADAMARD MATRICES 51 The problem comes of course from the assumption Bi,Cj ≥ 0, which is quite a subtle one. In what follows we restrict attention to the case p1 = 2. Here we have: Theorem 3.27. For q = 2p with p ≥ 3 prime, P2 equals the probability for a random walk on Zp to end up on the diagonal, i.e. at a position of type (t, . . . , t), with t ∈ Z. Proof. According to Proposition 3.26, we must understand the matrices A ∈ M2×p(N) which decompose as Aij = Bi + Cj, with Bi,Cj ≥ 0. But this is an easy task, because depending on A11 vs. A21 we have 3 types of solutions, as follows: a . . . a a . . . a a + t . . . a + t 1 p , 1 p , 1 p a1 . . . ap a1 + t . . . ap + t a1 . . . ap Here ai ≥ 0 and t ≥ 1. Now since cases 2,3 contribute in the same way, we obtain: 1 X N P2 = N (2p) a1, a1, . . . , ap, ap 2Σai=N 2 X X N + N (2p) a1, a1 + t, . . . , ap, ap + t t≥1 2Σai+pt=N We can write this formula in a more compact way, as follows: 1 X X N P2 = N (2p) a1, a1 + |t|, . . . , ap, ap + |t| t∈Z 2Σai+p|t|=N 1 Now since the sum on the right, when rescaled by (2p)N , is exactly the probability for p a random walk on Z to end up at (t, . . . , t), this gives the result. According to the above result we have P = P P (t), where P (t) with t ∈ is the 2 t∈Z 2 2 Z probability for a random walk on Zp to end up at (t, . . . , t). Observe that, by using Proposition 3.24 above with s, p, n equal respectively to p, 2,N/2, we obtain: 2 (0) 1 N X N/2 P2 = N (2p) N/2 a1, . . . , ap a1+...+ap=N/2 r 2 r pp ' × πN 2p−1(πN)p−1 r p p = 2 2πN (t) Regarding now the probability P2 of ending up at (t, . . . , t), in principle for small t this can be estimated by using a modification of the method in [81]. However, it is not clear on how to compute the full diagonal return probability in Theorem 3.27. 52 TEO BANICA It is possible to establish a few more results in this direction, and we refer here to [7]. However, the main question remains that of adapting the methods in [66] to the root of unity case. As a preliminary observation here, also from [7], we have: Theorem 3.28. The probability PM for a random H ∈ MM×N (Zq) to be partial Butson equals the probability for a length N random walk with increments drawn from n o M E = (eie¯j)i Proof. Indeed, with T (e) = (eie¯j)i 4. Geometry, defect In this section and in the next two ones we discuss various geometric aspects of the complex Hadamard matrices. Let us recall that the complex Hadamard manifold appears as an intersection of smooth real algebraic manifolds, as follows: √ XN = MN (T) ∩ NUN This intersection is very far from being smooth. Given a point H ∈ XN , the problem is that of understanding the structure of XN around H, which is often singular. There are several ways of discussing this question, a quite straightforward approach, going back to the work in [62], and then in [72], [88], being via the 1-parameter defor- mations of the complex Hadamard matrices. In what follows we will use an equivalent approach, of more real algebraic geometry flavor, developed in [5], [6]. We denote by Xp an unspecified neighborhood of a point in a manifold, p ∈ X. Also, r it r itr for q ∈ T1, meaning that q ∈ T is close to 1, we define q with r ∈ R by (e ) = e . With these conventions, we have the following result: Proposition 4.1. For H ∈ XN and A ∈ MN (R), the following are equivalent: q Aij (1) Hij = Hijq is an Hadamard matrix, for any q ∈ T1. P ¯ Aik−Ajk (2) k HikHjkq = 0, for any i 6= j and any q ∈ T1. P ¯ (3) k HikHjkϕ(Aik − Ajk) = 0, for any i 6= j and any ϕ : R → C. P ¯ r (4) r HikHjk = 0 for any i 6= j and r ∈ , where E = {k|Aik − Ajk = r}. k∈Eij R ij Proof. These equivalences are all elementary, and can be proved as follows: (1) ⇐⇒ (2) Indeed, the scalar products between the rows of Hq are: q q X Aik ¯ Ajk X ¯ Aik−Ajk < Hi ,Hj >= Hikq Hjkq¯ = HikHjkq k k (2) =⇒ (4) This follows from the following formula, and from the fact that the power functions {qr|r ∈ R} over the unit circle T are linearly independent: X ¯ Aik−Ajk X r X ¯ HikHjkq = q HikHjk r k r∈R k∈Eij (4) =⇒ (3) This follows from the following formula: X ¯ X X ¯ HikHjkϕ(Aik − Ajk) = ϕ(r) HikHjk r k r∈R k∈Eij r (3) =⇒ (2) This simply follows by taking ϕ(r) = q . Observe that in the above statement the condition (4) is purely combinatorial. In order to understand the above deformations, which are “affine” in a certain sense, it is convenient to enlarge the attention to all types of deformations. 54 TEO BANICA We keep using the neighborhood notation Xp introduced above, and we consider func- tions of type f : Xp → Yq, which by definition satisfy f(p) = q. With these conventions, let us introduce the following notions: Definition 4.2. Let H ∈ MN (C) be a complex Hadamard matrix. (1) A deformation of H is a smooth function f : T1 → (XN )H . Aij (2) The deformation is called “affine” if fij(q) = Hijq , with A ∈ MN (R). ai+bj N (3) We call “trivial” the deformations of type fij(q) = Hijq , with a, b ∈ R . it iAij t Here the adjective “affine” comes from fij(e ) = Hije , because the function t → Aijt which produces the exponent is indeed affine. As for the adjective “trivial”, this comes ai+bj from the fact that f(q) = (Hijq )ij is obtained from H by multiplying the rows and columns by certain numbers in T, so it is automatically Hadamard. The basic example of an affine deformation comes from the Dit¸˘adeformations H ⊗Q K, by taking all parameters qij ∈ T to be powers of q ∈ T. As an example, here are the exponent matrices coming from the left and right Dit¸˘adeformations of F2 ⊗ F2: a a b b a b a b c c d d a b a b Al = Ar = a a b b c d c d c c d d c d c d In order to investigate the above types of deformations, we will use the corresponding tangent vectors. So, let us recall that the manifold XN is given by: √ XN = MN (T) ∩ NUN This observation leads to the following definition, where in the first part we denote by TpX the tangent space to a point in a smooth manifold, p ∈ X: Definition 4.3. Associated to a point H ∈ X are the following objects: N √ (1) The enveloping tangent space: TeH XN = TH MN (T) ∩ TH NUN . (2) The tangent cone TH XN : the set of tangent vectors to the deformations of H. ◦ (3) The affine tangent cone TH XN : same as above, using affine deformations only. × (4) The trivial tangent cone TH XN : as above, using trivial deformations only. × ◦ Observe that TeH XN ,TH XN are real linear spaces, and that TH XN ,TH XN are two-sided cones, in the sense that they satisfy the following condition: λ ∈ R,A ∈ T =⇒ λA ∈ T Observe also that we have inclusions of cones, as follows: × ◦ TH XN ⊂ TH XN ⊂ TH XN ⊂ TeH XN In more algebraic terms now, these various tangent cones are best described by the corresponding matrices, and we have here the following result: HADAMARD MATRICES 55 × ◦ Theorem 4.4. The cones TH XN ⊂ TH XN ⊂ TH XN ⊂ TeH XN are as follows: (1) TeH XN can be identified with the linear space formed by the matrices A ∈ MN (R) P ¯ satisfying k HikHjk(Aik − Ajk) = 0, for any i, j. 0 (2) TH XN consists of those matrices A ∈ MN (R) appearing as Aij = gij(0), where P ¯ i(gik(t)−gjk(t)) g : MN (R)0 → MN (R)0 satisfies k HikHjke = 0 for any i, j. ◦ P ¯ Aik−Ajk (3) TH XN is formed by the matrices A ∈ MN (R) satisfying k HikHjkq = 0, for any i 6= j and any q ∈ T. × (4) TH XN is formed by the matrices A ∈ MN (R) which are of the form Aij = ai + bj, for certain vectors a, b ∈ RN . Proof. All these assertions can be deduced by using basic differential geometry: (1) This result is well-known, the idea being as follows. First, MN (T) is defined by the 2 algebraic relations |Hij| = 1, and with Hij = Xij + iYij we have: 2 2 2 ˙ ˙ d|Hij| = d(Xij + Yij) = 2(XijXij + YijYij) P ˙ ˙ Now since an arbitrary vector ξ ∈ TH MN (C), written as ξ = ij αijXij +βijYij, belongs 2 to TH MN (T) if and only if < ξ, d|Hij| >= 0 for any i, j, we obtain: ( ) X ˙ ˙ TH MN (T) = Aij(YijXij − XijYij) Aij ∈ R ij √ We also know that NUN is defined by the algebraic relations < Hi,Hj >= Nδij, where H1,...,HN are the rows of H. The relations < Hi,Hi >= N being automatic for the matrices H ∈ MN (T), if for i 6= j we let Lij =< Hi,Hj >, then we have: n ˙ o TeH CN = ξ ∈ TH MN (T)| < ξ, Lij >= 0, ∀i 6= j On the other hand, differentiating the formula of Lij gives: ˙ X ˙ ˙ ˙ ˙ Lij = (Xik + iYik)(Xjk − iYjk) + (Xjk − iYjk)(Xik + iYik) k Now if we pick ξ ∈ TH MN (T), written as above in terms of A ∈ MN (R), we obtain: ˙ X ¯ < ξ, Lij >= i HikHjk(Aik − Ajk) k Thus we have reached to the description of TeH XN in the statement. it igij (t) (2) Pick an arbitrary deformation, and write it as fij(e ) = Hije . Observe first that the Hadamard condition corresponds to the equations in the statement, namely: X ¯ i(gik(t)−gjk(t)) HikHjke = 0 k 56 TEO BANICA Observe also that by differentiating this formula at t = 0, we obtain: X ¯ 0 0 HikHjk(gik(0) − gjk(0)) = 0 k 0 Thus the matrix Aij = gij(0) belongs indeed to TeH XN , so we obtain in this way a certain map TH XN → TeH XN . In order to check that this map is indeed the correct one, we have to verify that, for any i, j, the tangent vector to our deformation is given by: 0 ˙ ˙ ξij = gij(0)(YijXij − XijYij) But this latter verification is just a one-variable problem. So, by dropping all i, j indices, which is the same as assuming N = 1, we have to check that for any point H ∈ T, written H = X + iY , the tangent vector to the deformation f(eit) = Heig(t) is: ξ = g0(0)(Y X˙ − XY˙ ) But this is clear, because the unit tangent vector at H ∈ T is η = −i(Y X˙ − XY˙ ), and ig(t) 0 0 its coefficient coming from the deformation is (e )|t=0 = −ig (0). (3) Observe first that by taking the derivative at q = 1 of the condition (2) in Propo- sition 4.1, of just by using the condition (3) there with the function ϕ(r) = r, we get: X ¯ HikHjkϕ(Aik − Ajk) = 0 k ◦ Thus we have a map TH XN → TeH XN , and the fact that is map is indeed the correct one comes for instance from the computation in (2), with gij(t) = Aijt. (4) Observe first that the Hadamard matrix condition is satisfied: X ¯ Aik−Ajk ai−aj X ¯ HikHjkq = q HikHjk = δij k k × As for the fact that TH XN is indeed the space in the statement, this is clear. Let ZN ⊂ XN be the real algebraic manifold formed by all the dephased N × N complex Hadamard matrices. Observe that we have a quotient map XN → ZN , obtained by dephasing. With this notation, we have the following refinement of (4) above: Proposition 4.5. We have a direct sum decomposition of cones ◦ × ◦ TH XN = TH XN ⊕ TH ZN where at right we have the affine tangent cone to the dephased manifold XN → ZN . ◦ Proof. If we denote by MN (R) the set of matrices having 0 outside the first row and column, we have a direct sum decomposition, as follows: ◦ ◦ ◦ TeH XN = MN (R) ⊕ TeH ZN Now by looking at the affine cones, and using Theorem 4.4, this gives the result. HADAMARD MATRICES 57 Summarizing, we have so far a number of theoretical results about the tangent cones TH XN that we are interested in, and their versions coming from√ the trivial and affine deformations, and from the intersection formula XN = MN (T) ∩ NUN as well. In practice now, passed a few special cases where all these cones collapse to the trivial × cone TN XN , which by Proposition 4.5 means that the image of H ∈ XN must be isolated in the dephased manifold XN → ZN , things are quite difficult to compute. However, as a concrete numerical invariant arising from all this, which can be effectively computed in many cases of interest, we have, following [88]: Definition 4.6. The real dimension d(H) of the enveloping tangent space √ TeH XN = TH MN (T) ∩ TH NUN is called undephased defect of a complex Hadamard matrix H ∈ XN . In view of Proposition 4.5, it is sometimes convenient to replace d(H) by the related quantity d0(H) = d(H) − 2N + 1, called dephased defect of H. See [88]. In what follows we will rather use d(H) as defined above, and simply call it “defect” of H. We already know, from Theorem 4.4, what is the precise geometric meaning of the defect, and how to compute it. Let us record again these results, that we will use many times in what follows, in a slightly different form, closer to the spirit of [88]: Theorem 4.7. The defect d(H) is the real dimension of the linear space ( ) X ¯ TeH XN = A ∈ MN (R) HikHjk(Aik − Ajk) = 0, ∀i, j k q Aij and the elements of this space are those making Hij = Hijq Hadamard at order 1. Proof. Here the first assertion is something that we already know, from Theorem 4.4 (1), and the second assertion follows either from Theorem 4.4 and its proof, or directly from the definition of the enveloping tangent space TeH XN , as used in Definition 4.6. Here are a few basic properties of the defect: Proposition 4.8. Let H ∈ XN be a complex Hadamard matrix. (1) If H ' He then d(H) = d(He). (2) We have 2N − 1 ≤ d(H) ≤ N 2. (3) If d(H) = 2N − 1, the image of H in the dephased manifold XN → ZN is isolated. Proof. All these results are elementary, the proof being as follows: (1) If we let Kij = aibjHij with |ai| = |bj| = 1 be a trivial deformation of our matrix H, the equations for the enveloping tangent space for K are: X ¯ ¯ aibkHika¯jbkHjk(Aik − Ajk) = 0 k 58 TEO BANICA By simplifying we obtain the equations for H, so d(H) is invariant under trivial defor- mations. Since d(H) is invariant as well by permuting rows or columns, we are done. × × (2) Consider the inclusions TH XN ⊂ TH XN ⊂ TeH XN . Since dim(TH XN ) = 2N − 1, the inequality at left holds indeed. As for the inequality at right, this is clear. × (3) If d(H) = 2N − 1 then TH XN = TH XN , so any deformation of H is trivial. Thus the image of H in the quotient manifold XN → ZN is indeed isolated, as stated. Let us discuss now the computation of the defect for the most basic examples of complex Hadamard matrices that we know, namely the real ones, and the Fourier ones. In order to deal with the real case, it is convenient to modify the general formula from Theorem 4.7 above, via a change of variables, as follows: Proposition 4.9. We have a linear space isomorphism as follows, n o ∗ ¯ TeH XN ' E ∈ MN (C) E = E , (EH)ijHij ∈ R, ∀i, j the correspondences A → E and E → A being given by the formulae X ¯ ¯ Eij = HikHjkAik ,Aij = (EH)ijHij k with A ∈ TeH XN being the usual components, from Theorem 4.7 above. ∗ Proof. Given a matrix A ∈ MN (C), if we set Rij = AijHij and E = RH , the correspon- dence A → R → E is then bijective onto MN (C), and we have: X ¯ Eij = HikHjkAik k In terms of these new variables, the equations in Theorem 4.7 become: ¯ Eij = Eji Thus, when taking into account these conditions, we are simply left with the conditions ¯ Aij ∈ R. But these correspond to the conditions (EH)ijHij ∈ R, as claimed. With the above result in hand, we can now compute the defect of the real Hadamard matrices. The result here, from [86], is as follows: Theorem 4.10. For any real Hadamard matrix H ∈ MN (±1) we have symm TeH XN ' MN (R) and so the corresponding defect is d(H) = N(N + 1)/2. ¯ Proof. We use Proposition 4.9. Since H is now real the condition (EH)ijHij ∈ R there simply tells us that E must be real, and this gives the result. HADAMARD MATRICES 59 We should mention that the above result, as well as the whole basic theory of the tangent cones and defect, can be extended in a quite straightforward way to the case of the partial Hadamard matrices [22]. We will be back to this, later on. Let us discuss now the computation of the defect of the Fourier matrix FG. The main idea here goes back to [62], with some supplementary contributions from [72], the main formula, in the cyclic group case, was obtained in [88], the extension to the general case was done in [5], and the corresponding deformations were studied in [73]. As a first result on this subject, we have, following [88]: Theorem 4.11. For a Fourier matrix F = FG, the matrices A ∈ TeF XN , with N = |G|, ∗ are those of the form A = PF , with P ∈ MN (C) satisfying ¯ Pij = Pi+j,j = Pi,−j where the indices i, j are by definition taken in the group G. Proof. We use the system of equations in Theorem 4.7, namely: X ¯ FikFjk(Aik − Ajk) = 0 k By decomposing our finite abelian group as G = ZN1 × ... × ZNr we can assume 2πi/k F = FN1 ⊗ ... ⊗ FNr , so that with wk = e we have: i1j1 irjr Fi1...ir,j1...jr = (wN1 ) ... (wNr ) 2πi/N With N = N1 ...Nr and w = e , we obtain: i1j1 +...+ irjr N N1 Nr Fi1...ir,j1...jr = w Thus the matrix of our system is given by: (i1−j1)k1 +...+ (ir−jr)kr N ¯ N1 Nr Fi1...ir,k1...kr Fj1...jr,k1...kr = w Now by plugging in a multi-indexed matrix A, our system becomes: X (i1−j1)k1 +...+ (ir−jr)kr N N1 Nr w (Ai1...ir,k1...kr − Aj1...jr,k1...kr ) = 0 k1...kr Now observe that in the above formula we have in fact two matrix multiplications, so our system can be simply written as: (AF )i1...ir,i1−j1...ir−jr − (AF )j1...jr,i1−j1...ir−jr = 0 Now recall that our indices have a “cyclic” meaning, so they belong in fact to the group G. So, with P = AF , and by using multi-indices, our system is simply: Pi,i−j = Pj,i−j With i = I + J, j = I we obtain the condition PI+J,J = PIJ in the statement. 60 TEO BANICA ∗ ˜ ¯ In addition, A = PF must be a real matrix. But, if we set Pij = Pi,−j, we have: X (PF ∗) = P¯ F i1...ir,j1...jr i1...ir,k1...kr j1...jr,k1...kr k1...kr X ˜ ∗ = Pi1...ir,−k1...−kr (F )−k1...−kr,j1...jr k1...kr ˜ ∗ = (PF )i1...ir,j1...jr Thus we have PF ∗ = PF˜ ∗, so the fact that the matrix PF ∗ is real, which means by definition that we have PF ∗ = PF ∗, can be reformulated as PF˜ ∗ = PF ∗, and hence as ˜ ¯ P = P . So, we obtain the conditions Pij = Pi,−j in the statement. We can now compute the defect, and we are led to the following formula: Theorem 4.12. The defect of a Fourier matrix FG is given by X |G| d(F ) = G ord(g) g∈G and equals as well the number of 1 entries of the matrix FG. ∗ Proof. According to the formula A = PF from Theorem 4.11, the defect d(FG) is the dimension of the real vector space formed by the matrices P ∈ MN (C) satisfying: ¯ Pij = Pi+j,j = Pi,−j Here, and in what follows, the various indices i, j, . . . will be taken in G. Now the point is that, in terms of the columns of our matrix P , the above conditions are: (1) The entries of the j-th column of P , say C, must satisfy Ci = Ci+j. (2) The (−j)-th column of P must be conjugate to the j-th column of P . Thus, in order to count the above matrices P , we can basically fill the columns one by one, by taking into account the above conditions. In order to do so, consider the subgroup G2 = {j ∈ G|2j = 0}, and then write G as a disjoint union, as follows: G = G2 t X t (−X) With this notation, the algorithm is as follows. First, for any j ∈ G2 we must fill the j-th column of P with real numbers, according to the periodicity rule Ci = Ci+j. Then, for any j ∈ X we must fill the j-th column of P with complex numbers, according to the same periodicity rule Ci = Ci+j. And finally, once this is done, for any j ∈ X we just have to set the (−j)-th column of P to be the conjugate of the j-th column. So, let us compute the number of choices for filling these columns. Our claim is that, when uniformly distributing the choices for the j-th and (−j)-th columns, for j∈ / G2, there are exactly [G :< j >] choices for the j-th column, for any j. Indeed: (1) For the j-th column with j ∈ G2 we must simply pick N real numbers subject to the condition Ci = Ci+j for any i, so we have indeed [G :< j >] such choices. HADAMARD MATRICES 61 (2) For filling the j-th and (−j)-th column, with j∈ / G2, we must pick N complex numbers subject to the condition Ci = Ci+j for any i. Now since there are [G :< j >] choices for these numbers, so a total of 2[G :< j >] choices for their real and imaginary parts, on average over j, −j we have [G :< j >] choices, and we are done again. Summarizing, the dimension of the vector space formed by the matrices P , which is equal to the number of choices for the real and imaginary parts of the entries of P , is: X d(FG) = [G :< j >] j∈G But this is exactly the number in the statement. Regarding now the second assertion, according to the abstract definition of the Fourier matrix FG, from Theorem 2.8 above, the number of 1 entries of FG is given by: n o #(1 ∈ FG) = # (g, χ) ∈ G × Gb χ(g) = 1 n o X = # χ ∈ Gb χ(g) = 1 g∈G X |G| = ord(g) g∈G Thus, the second assertion follows from the first one. Let us finish now the work, and explicitely compute the defect of FG. It is convenient to consider the following quantity, which behaves better: X 1 δ(G) = ord(g) g∈G a As a first example, consider a cyclic group G = ZN , with N = p power of a prime. The count here is very simple, over sets of elements having a given order: −1 2 −2 a a−1 −1 δ(Zpa ) = 1 + (p − 1)p + (p − p)p + ... + (p − p )p a = 1 + a − p In order to extend this kind of count to the general abelian case, we use two ingredients. First is the following result, which splits the computation over isotypic components: Proposition 4.13. For any finite groups G, H we have: δ(G × H) ≥ δ(G)δ(H) In addition, if (|G|, |H|) = 1, we have equality. 62 TEO BANICA Proof. Indeed, we have the following estimate: X 1 δ(G × H) = ord(g, h) gh X 1 = [ord(g), ord(h)] gh X 1 ≥ ord(g) · ord(h) gh = δ(G)δ(H) Now in the case (|G|, |H|) = 1, the least common multiple appearing on the right becomes a product, [ord(g), ord(h)] = ord(g) · ord(h), so we have equality. We deduce from this that we have the following result: Proposition 4.14. For a finite abelian group G we have Y δ(G) = δ(Gp) p where Gp with G = ×pGp are the isotypic components of G. Proof. This is clear from Proposition 4.13, because the order of Gp is a power of p. As an illustration, we can recover in this way the defect computation in [89]: Theorem 4.15. The defect of a usual Fourier matrix FN is given by s Y ai d(F ) = N 1 + a − N i p i=1 i a1 as where N = p1 . . . ps is the decomposition of N into prime factors. Proof. The underlying group here is the cyclic group G = ZN , whose isotypic components are the cyclic groups Gp = ai . By applying now Proposition 4.14, and by using the i Zpi computation for cyclic p-groups performed before Proposition 4.13, we obtain: s Y −1 d(FN ) = N 1 + pi (pi − 1)ai i=1 But this is exactly the formula in the statement. Now back to the general case, where we have an arbitrary Fourier matrix FG, we will need, as a second ingredient for our computation, the following result: HADAMARD MATRICES 63 Proposition 4.16. For the p-groups, the quantities n o k ck = # g ∈ G ord(g) ≤ p are multiplicative, in the sense that ck(G × H) = ck(G)ck(H). Proof. Indeed, for a product of p-groups we have: n o k ck(G × H) = # (g, h) ord(g, h) ≤ p n o k k = # (g, h) ord(g) ≤ p , ord(h) ≤ p n o n o k k = # g ord(g) ≤ p # h ord(h) ≤ p We recognize at right ck(G)ck(H), and we are done. Let us compute now δ in the general isotypic case: Proposition 4.17. For G = Zpa1 × ... × Zpar with a1 ≤ a2 ≤ ... ≤ ar we have r X (r−k)ak−1+(a1+...+ak−1)−1 r−k+1 δ(G) = 1 + p (p − 1)[ak − ak−1]pr−k k=1 2 a−1 with the convention a0 = 0, and by using the notation [a]q = 1 + q + q + ... + q . Proof. First, in terms of the numbers ck, we have: X ck − ck−1 δ(G) = 1 + pk k≥1 min(k,a) In the case of a cyclic group G = Zpa we have ck = p . Thus, in the general isotypic case G = Zpa1 × ... × Zpar we have: min(k,a1) min(k,ar) min(k,a1)+...+min(k,ar) ck = p . . . p = p Now observe that the exponent on the right is a piecewise linear function of k. More precisely, by assuming a1 ≤ a2 ≤ ... ≤ ar as in the statement, the exponent is linear on each of the intervals [0, a1], [a1, a2],..., [ar−1, ar]. So, the quantity δ(G) to be computed will be 1 plus the sum of 2r geometric progressions, 2 for each interval. In practice now, the numbers ck are as follows: r 2r ra1 c0 = 1, c1 = p , c2 = p , . . . , ca1 = p , a1+(r−1)(a1+1) a1+(r−1)(a1+2) a1+(r−1)a2 ca1+1 = p , ca1+2 = p , . . . , ca2 = p , a1+a2+(r−2)(a2+1) a1+a2+(r−2)(a2+2) a1+a2+(r−2)a3 ca2+1 = p , ca2+2 = p , . . . , ca3 = p , ...... a1+...+ar−1+(ar−1+1) a1+...+ar−1+(ar−1+2) a1+...+ar car−1+1 = p , car−1+2 = p , . . . , car = p 64 TEO BANICA Now by separating the positive and negative terms in the above formula of δ(G), we have indeed 2r geometric progressions to be summed, as follows: δ(G) = 1 + (pr−1 + p2r−2 + p3r−3 + ... + pa1r−a1 ) −(p−1 + pr−2 + p2r−3 + ... + p(a1−1)r−a1 ) +(p(r−1)(a1+1)−1 + p(r−1)(a1+2)−2 + ... + pa1+(r−2)a2 ) −(pa1r−a1−1 + p(r−1)(a1+1)−2 + ... + pa1+(r−1)(a2−1)−a2 ) + ... +(pa1+...+ar−1 + pa1+...+ar−1 + ... + pa1+...+ar−1 ) −(pa1+...+ar−1−1 + pa1+...+ar−1−1 + ... + pa1+...+ar−1−1) Now by performing all the sums, we obtain: p(r−1)a1 − 1 δ(G) = 1 + p−1(pr − 1) pr−1 − 1 p(r−2)(a2−a1) − 1 +p(r−2)a1+(a1−1)(pr−1 − 1) pr−2 − 1 p(r−3)(a3−a2) − 1 +p(r−3)a2+(a1+a2−1)(pr−2 − 1) pr−3 − 1 + ... a1+...+ar−1−1 +p (p − 1)(ar − ar−1) By looking now at the general term, we get the formula in the statement. Let us go back now to the general defect formula in Theorem 4.12. By putting it together with the various results above, we obtain: Theorem 4.18. For a finite abelian group G, decomposed as G = ×pGp, we have r ! Y X (r−k)ak−1+(a1+...+ak−1)−1 r−k+1 d(FG) = |G| 1 + p (p − 1)[ak − ak−1]pr−k p k=1 where a0 = 0 and a1 ≤ a2 ≤ ... ≤ ar are such that Gp = Zpa1 × ... × Zpar . Proof. Indeed, we know from Theorem 4.12 that we have d(FG) = |G|δ(G), and the result follows from Proposition 4.14 and Proposition 4.17. HADAMARD MATRICES 65 As a first illustration, we can recover in this way the formula in Theorem 4.15. Assuming a1 as that N = p1 . . . ps is the decomposition of N into prime factors, we have: s Y −1 d(FN ) = N 1 + pi (pi − 1)ai i=1 s Y ai = N 1 + a − i p i=1 i As a second illustration, for the group G = Zpa1 × Zpa2 with a1 ≤ a2 we obtain: a1+a2 −1 2 a1−1 d(FG) = p (1 + p (p − 1)[a1]p + p (p − 1)(a2 − a1)) pa1 − 1 = pa1+a2−1(p + (p2 − 1) + pa1 (p − 1)(a − a )) p − 1 2 1 a1+a2−1 a1 a1 = p (p + (p + 1)(p − 1) + p (p − 1)(a2 − a1)) Finally, let us mention that for general non-abelian groups, there does not seem to be any reasonable algebraic formula for the quantity δ(G). As an example, consider the dihedral group DN , consisting of N symmetries and N rotations. We have: N δ(D ) = + δ( ) N 2 ZN Q −1 Now by remembering the formula for ZN , namely δ(ZN ) = (1 + pi (pi − 1)ai), it is quite clear that the N/2 factor can not be incorporated in any nice way. See [5]. Let us prove now, following the paper of Nicoara and White [73], that for the Fourier matrices the defect is “attained”, in the sense that the deformations at order 0 are true deformations, at order ∞. This is something quite surprising, and non-trivial. Let us begin with some generalities. We first recall that we have: Proposition 4.19. The unitary matrices U ∈ UN around 1 are of the form U = eA with A being an antihermitian matrix, A = −A∗, around 0. Proof. This is something well-known. Indeed, assuming that a matrix A is antihermitian, A = −A∗, the matrix U = eA follows to be unitary: UU ∗ = eA(eA)∗ = eAeA∗ = eAe−A = 1 As for the converse, this follows either by using a dimension argument, which shows that the space of antihermitian matrices is the correct one, or by diagonalizing U. Now back to the Hadamard matrices, we will need to rewrite a part of the basic theory of the defect, using deformations of type t → UtH. First, we have: 66 TEO BANICA Theorem 4.20. Assume that H ∈ MN (C) is Hadamard, let A ∈ MN (C) be antihermitian, and consider the matrix UH, where U = etA, with t ∈ R. P ¯ tA −tA (1) UH is Hadamard when | rs HrqHsq(e )pr(e )sp| = 1, for any p, q. (2) UH is Hadamard at order 0 when |(AH)pq| = 1, for any p, q. Proof. We already know that UH is unitary, so we must find the conditions which guar- antee that we have UH ∈ MN (T), in general, and then at order 0. (1) We have the following computation, valid for any unitary U: 2 |(UH)pq| = (UH)pq(UH)pq ∗ ∗ = (UH)pq(H U )qp X ∗ ∗ = UprHrq(H )qs(U )sp rs X ¯ ¯ = HrqHsqUprUps rs Now with U = etA as in the statement, we obtain: tA 2 X ¯ tA −tA |(e H)pq| = HrqHsq(e )pr(e )sp rs Thus, we are led to the conclusion in the statement. (2) The derivative of the function computed above, taken at 0, is as follows: tA 2 ∂|(e H)pq| X ¯ tA tA = HrqHsq(e A)pr(−e A)sp |t=0 ∂t |t=0 rs X ¯ = HrqHsqApr(−A)sp rs X X ∗ ∗ = AprHrq (H )qs(A )sp r s ∗ ∗ = (AH)pq(H A )qp 2 = |(AH)pq| Thus, we obtain the conclusion in the statement. In the Fourier matrix case we can go beyond this, and we have: Proposition 4.21. Given a Fourier matrix FG ∈ MG(C), and an antihermitian matrix tA A ∈ MG(C), the matrix UFG, where U = e with t ∈ R, is Hadamard when X X tm X m X Ak (−A)l = δ m! l p,s+n sp n0 s m k+l=m s for any p, with the indices being k, l, m ∈ N, and n, p, s ∈ G. HADAMARD MATRICES 67 Proof. According to the formula in the proof of Theorem 4.20 (1), we have: 2 X tA −tA |(UFG)pq| = (FG)rq(FG)sq(e )pr(e )sp rs X tA −tA = < r, q >< −s, q > (e )pr(e )sp rs X tA −tA = < r − s, q > (e )pr(e )sp rs By setting n = r − s, can write this formula in the following way: 2 X tA −tA |(UFG)pq| = < n, q > (e )p,s+n(e )sp ns X X tA −tA = < n, q > (e )p,s+n(e )sp n s Since this quantity must be 1 for any q, we must have: X tA −tA (e )p,s+n(e )sp = δn0 s On the other hand, we have the following computation: X tA −tA (e )p,s+n(e )sp s k l X X (tA)p,s+n (−tA)sp = · k! l! s kl X X 1 X = (tA)k (−tA)l k!l! p,s+n sp s kl s X X tk+l X = Ak (−A)l k!l! p,s+n sp s kl s X X X 1 X = tm Ak (−A)l k!l! p,s+n sp s m k+l=m s X X tm X m X = Ak (−A)l m! l p,s+n sp s m k+l=m s Thus, we obtain the conclusion in the statement. Following [73], let us construct now the deformations. The result is as follows: 68 TEO BANICA Theorem 4.22. Let G be a finite abelian group, and for any g, h ∈ G, let us set: ( 1 if ∃k ∈ , p = hkg, q = hk+1g B = N pq 0 otherwise When (g, h) ∈ G2 range in suitable cosets, the unitary matrices it(B+Bt) t(B−Bt) e FG , e FG are both Hadamard, and make the defect of FG to be attained. Proof. The proof of this result, from [73], is quite long and technical, based on the Fourier computation from Proposition 4.21 above, the idea being as follows: (1) First of all, an elementary algebraic study shows that when (g, h) ∈ G2 range in some suitable cosets, coming from the proof of Theorem 4.12, the various matrices B = Bgh constructed above are distinct, the matrices A = i(B + Bt) and A0 = B − Bt are linearly independent, and the number of such matrices equals the defect of FG. (2) It is also standard to check that each B = (Bpq) is a partial isometry, and that Bk,B∗k are given by simple formulae. With this ingredients in hand, the Hadamard property follows from the Fourier computation from the proof of Proposition 4.21. Indeed, we can compute the exponentials there, and eventually use the binomial formula. (3) Finally, the matrices in the statement can be shown to be non-equivalent, and this is something more technical, for which we refer to [73]. With this last ingredient in hand, a comparison with Theorem 4.12 shows that the defect of FG is indeed attained, in the sense that all order 0 deformations are actually true deformations. See [73]. The above result is something quite surprising, which came a long time after the original defect paper [88], and even more time after the early computations in [62]. Let us also mention that [73] was written in terms of subfactor-theoretic commuting squares, with a larger class of squares actually under investigation. We will discuss the relation between Hadamard matrices and commuting squares in section 11 below. HADAMARD MATRICES 69 5. Special matrices We have seen in the previous section that the defect theory from [88] can be successfully applied to the real Hadamard matrices, and to the Fourier matrices. We discuss here a number of more specialized aspects, regarding the tensor products, the Dit¸˘adeformations of such tensor products, the Butson and the regular matrices, the master Hadamard matrices, the McNulty-Weigert matrices, and finally the partial Hadamard matrix case, following [2], [5], [6], [22], [69], [88], [89]. Let us begin with some generalities. The standard defect equations are those in Theo- rem 4.7, naturally coming from the computations in the proof of Theorem 4.4: ( ) X d(H) = dim A ∈ M ( ) H H¯ (A − A ) = 0, ∀i, j R N R ik jk ik jk k However, we have seen that for various concrete questions, some manipulations on these equations are needed. To be more precise, the study in the real case was based on the P ¯ transformation Eij = k HikHjkAik from Proposition 4.9, the study in the Fourier matrix case was based on the transformation P = AFG from Theorem 4.11, and the fine study in the Fourier matrix case was based on the t → etAH method from Proposition 4.21. In short, each type of complex Hadamard matrix seems to require its own general theory, and defect manipulations, and there is no way of escaping from this. In view of this phenomenon, let us first present some further manipulations on the defect equations, which are all quite natural, and potentially useful. First, we have: Proposition 5.1. The defect d(H) is the corank of the matrix Re(H H¯ ) if i < j ib jb ¯ Yij,ab = (δia − δja) Im(HibHjb) if i > j ∗ if i = j where ∗ can be any quantity, its coefficient being 0 anyway. Proof. The matrix of the system defining the enveloping tangent space is: ¯ Xij,ab = (δia − δja)HibHjb However, since we are only looking for real solutions A ∈ MN (R), we have to take into account the real and imaginary parts. But this is not a problem, because the (ij) equation coincides with the (ji) one, that we can cut. More precisely, if we set Y as above, then we obtain precisely the original system. Thus the defect of H is the corank of Y . 70 TEO BANICA As an illustration, for the Fourier matrix FN we have the following formula, where e(i, j) ∈ {−1, 0, 1} is negative if i < j, null for i = j, and positive for i > j: 1 Y = (δ − δ )(w(i−j)b + e(i, j)w(j−i)b) ij,ab 2 ia ja Observe in particular that for the Fourier matrix F2 we have: 0 0 0 0 1 −1 −1 1 Y = −1 1 1 −1 0 0 0 0 Here the corank is 3, but, unfortunately, this cannot be seen on the characteristic polynomial, which is P (λ) = λ4. The problem is that our matrix, and more precisely its middle 2 × 2 block, is not diagonalizable. This phenomenon seems to hold in general. A second possible manipulation, which is of interest in connection with quantum groups and subfactor theory, concerns the reformulation of the defect in terms of the profile matrix of H. Indeed, it is known that both the quantum group and subfactor associated to H depend only on this profile matrix, and this will be explained in sections 10-11 below. We do not have an answer here, but our conjecture is as follows: Conjecture 5.2. The profile matrix of H, namely jb X ¯ ¯ Mia = HikHjkHakHbk k determines the enveloping tangent space TeH XN , or at least the defect d(H). All this is of course related to the general question on whether the associated quantum group or subfactor can “see” the defect, via various representation theory invariants. For a number of further speculations on all this, and on some related glow questions as well, in relation with the general theory in [45], [46], we refer to [5], [6], [8]. Let us get back now to our original goal here, namely computing the defect for various classes of special matrices. For the tensor products, we have the following result: Proposition 5.3. For a tensor product L = H ⊗ K we have d(L) ≥ d(H)d(K) coming from an inclusion of linear spaces TeH XM ⊗ TeK XN ⊂ TeLXMN . HADAMARD MATRICES 71 Proof. For a matrix A = B ⊗ C, we have the following formulae: X X ¯ X ¯ (H ⊗ K)ia,kc(H ⊗ K)jb,kcAia,kc = HikHjkBik KacKbcCac kc k c X X ¯ X ¯ (H ⊗ K)ia,kc(H ⊗ K)jb,kcAjb,kc = HikHjkBjk KacKbcCbc kc k c Now by assuming B ∈ TeH XM and C ∈ TeK XN , the two quantities on the right are equal. Thus we have indeed A ∈ TeLXMN , and we are done. Observe that we do not have equality in the tensor product estimate, even in very simple cases. For instance if we consider two Fourier matrices F2, we obtain: 2 d(F2 ⊗ F2) = 10 > 9 = d(F2) In fact, besides the isotypic decomposition results from section 4 above, valid for the Fourier matrices, there does not seem to be anything conceptual on this subject. We will be back to this, however, in Theorem 5.6 below, with a slight advance on all this. Let us discuss now the Dit¸˘adeformations. Here the study is even more difficult, and we basically have just one result, when the deformation matrix is as follows: Definition 5.4. A rectangular matrix Q ∈ MM×N (T) is called “dephased and elsewhere generic” if the entries on its first row and column are all equal to 1, and the remaining (M − 1)(N − 1) entries are algebrically independent over Q. Here the last condition takes of course into account the fact that the entries of Q themselves have modulus 1, the independence assumption being modulo this fact. With this convention made, we have the following result: Theorem 5.5. If H ∈ XM ,K ∈ XN are dephased, of Butson type, and Q ∈ MM×N (T) is dephased and elsewhere generic, then A = (Aia,kc) belongs to TeH⊗QK XMN iff ij ij ij ji ii Aac = Abc ,Aac = Aac , (Axy)xy ∈ TeK XN ij P ¯ hold for any a, b, c and i 6= j, where Aac = k HikHjkAia,kc. Proof. Consider the system for the enveloping tangent space, namely: X (H ⊗Q K)ia,kc(H ⊗Q K)jb,kc(Aia,kc − Ajb,kc) = 0 kc We have (H ⊗Q K)ia,jb = qibHijKab, and so our system is: X ¯ X ¯ qicq¯jcKacKbc HikHjk(Aia,kc − Ajb,kc) = 0 c k 72 TEO BANICA ij P ¯ Consider now the variables Aac = k HikHjkAia,kc in the statement. We have: ij X ¯ X ¯ Aac = HikHjkAia,kc = HjkHikAia,kc k k Thus, in terms of these variables, our system becomes simply: X ¯ ij ji qicq¯jcKacKbc(Aac − Abc) = 0 c More precisely, the above equations must hold for any i, j, a, b. By distinguishing now two cases, depending on whether i, j are equal or not, the situation is as follows: (1) Case i 6= j. In this case, let us look at the row vector of parameters, namely: (qicq¯jc)c = (1, qi1q¯j1, . . . , qiM q¯jM ) Now since Q was assumed to be dephased and elsewhere generic, and because of our assumption i 6= j, the entries of the above vector are linearly independent over Q¯ . But, since by linear algebra we can restrict attention to the computation of the solutions over ¯ ij ji Q, the i 6= j part of our system simply becomes Aac = Abc, for any a, b, c and any i 6= j. Now by making now a, b, c vary, we are led to the following equations: ij ij ij ji Aac = Abc,Aac = Aac, ∀a, b, c, i 6= j (2) Case i = j. In this case the parameters cancel, and our equations become: X ¯ ii ii KacKbc(Aac − Abc) = 0, ∀a, b, c, i c ii P On the other hand, we have Aac = k Aia,kc, and so our equations become: X ¯ ii ii KacKbc(Aac − Abc) = 0, ∀a, b, c, i c But these are precisely the equations for the space TeK XN , and we are done. Let us go back now to the usual tensor product situation, and look at the affine cones. ◦ The problem here is that of finding the biggest subcone of TH⊗K XMN , obtained by gluing ◦ ◦ TH XM ,TK XN . Our answer here, which takes into account the two “semi-trivial” cones coming from the left and right Dit¸˘adeformations, is as follows: ◦ ◦ Theorem 5.6. The cones TH XM = {B} and TK XN = {C} glue via the formulae Aia,jb = λBij + ψjCab + Xia + Yjb + Faj Aia,jb = φbBij + µCab + Xia + Yjb + Eib ◦ producing in this way two subcones of the affine cone TH⊗K XMN = {A}. HADAMARD MATRICES 73 Proof. Indeed, the idea is that Xia,Yjb are the trivial parameters, and that Eib,Faj are the Dit¸˘aparameters. In order to prove the result, we use the criterion in Theorem 4.4 (3) above. So, given a matrix A = (Aia,jb), consider the following quantity: X ¯ ¯ Aia,kc−Ajb,kc P = HikHjkKacKbcq kc Let us prove now the first statement, namely that for any choice of matrices B ∈ ◦ ◦ TH XM ,C ∈ TH XN and of parameters λ, ψj,Xia,Yjb,Faj, the first matrix A = (Aia,jb) ◦ constructed in the statement belongs indeed to TH⊗K XMN . We have: Aia,kc = λBik + ψkCac + Xia + Ykc + Fak Ajb,kc = λBjk + ψkCbc + Xjb + Ykc + Fbk Now by substracting, we obtain: Aia,kc − Ajb,kc = λ(Bik − Bjk) + ψk(Cac − Cbc) + (Xia − Xjb) + (Fak − Fbk) It follows that the above quantity P is given by: X ¯ ¯ λ(Bik−Bjk)+ψk(Cac−Cbc)+(Xia−Xjb)+(Fak−Fbk) P = HikHjkKacKbcq kc Xia−Xjb X ¯ Fak−Fbk λ(Bik−Bjk) X ¯ ψk Cac−Cbc = q HikHjkq q KacKbc(q ) k c Xia−Xja X ¯ λ Bik−Bjk = δabq HikHjk(q ) k = δabδij ◦ Thus Theorem 4.4 (3) applies and tells us that we have A ∈ TH⊗K XMN , as claimed. In the second case now, the proof is similar. First, we have: Aia,kc = φcBik + µCac + Xia + Ykc + Eic Ajb,kc = φcBjk + µCbc + Xjb + Ykc + Ejc Thus by substracting, we obtain: Aia,kc − Ajb,kc = φc(Bik − Bjk) + µ(Cac − Cbc) + (Xia − Xjb) + (Eic − Ejc) It follows that the above quantity P is given by: X ¯ ¯ φc(Bik−Bjk)+µ(Cac−Cbc)+(Xia−Xjb)+(Eic−Ejc) P = HikHjkKacKbcq kc Xia−Xjb X ¯ Eic−Ejc µ(Cac−Cbc) X ¯ φc Bik−Bjk = q KacKbcq q HikHjk(q ) c k Xia−Xib X ¯ µ Cac−Cbc = δijq KacKbc(q ) = δijδab c Thus Theorem 4.4 (3) applies again, and gives the result. 74 TEO BANICA We believe Theorem 5.6 above to be “optimal”, in the sense that nothing more can be ◦ said about the affine tangent space TH⊗K XMN , in the general case. See [6]. Let us discuss now some rationality questions, in relation with: Definition 5.7. The rational enveloping tangent space at H ∈ XN is [TeH XN ]Q = TeH XN ∩ MN (Q) and the dimension dQ(H) of this space is called rational defect of H. Observe that the first notion can be extended to all the tangent cones at H, and by using an arbitrary field K ⊂ C instead of Q. Indeed, we can set: ∗ ∗ [TH XN ]K = TH XN ∩ MN (K) However, in what follows we will be interested only in the objects constructed in Defi- nition 5.7. It follows from definitions that dQ(H) ≤ d(H), and we have: Conjecture 5.8. For a regular Hadamard matrix H ∈ MN (C) we have TeH CN = C · [TeH CN ]Q and so the defect equals the rational defect, dQ(H) = d(H). For the usual Fourier matrices FN , this definitely holds at N = p prime, because the minimal polynomial of w over Q is simply P = 1 + w + ... + wp−1. The case N = p2 has also a simple solution, coming from the fact that all the p × p blocks of our matrix A can be shown to coincide. In general, this method should probably work for N = pk, or even for any N ∈ N. So, our conjecture would be that this holds indeed for the Fourier matrices, usual or general, and more generally for the Butson matrices, and even more generally, modulo Conjecture 3.22 above, for the regular matrices. See [6]. Let us discuss now defect computations for a very interesting class of Hadamard ma- trices, namely the “master” ones, introduced in [2]: Definition 5.9. A master Hadamard matrix is an Hadamard matrix of the form nj Hij = λi P nj with λi ∈ T, nj ∈ R. The associated “master function” is f(z) = j z . imi iminj Observe that with λi = e we have Hij = e . The basic example of such a matrix is the Fourier matrix FN , having master function as follows: zN − 1 f(z) = z − 1 Observe that, in terms of f, the Hadamard condition on H is simply: λi f = Nδij λj HADAMARD MATRICES 75 These matrices were introduced in [2], the motivating remark there being the fact that the following operator defines a representation of the Temperley-Lieb algebra [92]: X ni−nj R = eij ⊗ Λ ij At the level of examples, the first observation, from [2], is that the standard 4 × 4 complex Hadamard matrices are, with 2 exceptions, master Hadamard matrices: Proposition 5.10. The following complex Hadamard matrix, with |q| = 1, 1 1 1 1 q 1 −1 1 −1 F = 2,2 1 q −1 −q 1 −q −1 q is a master Hadamard matrix, for any q 6= ±1. Proof. We use the exponentiation convention (eit)r = eitr for t ∈ [0, 2π) and r ∈ R. Since q2 6= 1, we can find k ∈ R such that q2k = −1, and so our matrix becomes: 10 11 12k 12k+1 0 1 2k 2k+1 q (−1) (−1) (−1) (−1) F = 0 1 2k 2k+1 2,2 q q q q (−q)0 (−q)1 (−q)2k (−q)2k+1 x y z t Now if we pick λ 6= 1 and write 1 = λ , −1 = λ , q = λ , −q = λ , we are done. We have the following generalization of Proposition 5.10, once again from [2]: Theorem 5.11. FM ⊗Q FN is master Hadamard, for any Q ∈ MM×N (T) of the form i(Npb+b) Qib = q 2πi/MNk where q = e with k ∈ N, and p0, . . . , pN−1 ∈ R. Proof. The main construction in [2] is, in terms of master functions, as follows: Nk f(z) = fM (z )fN (z) Here k ∈ N, and the functions on the right are by definition as follows: X Mri+i X Npa+a fM (z) = z , fN (z) = z i a i a 2πi/N Nk We use the eigenvalues λia = q w , where w = e , and where q = ν, where M Nk ν = 1. Observe that, according to f(z) = fM (z )fN (z), the exponents are: njb = Nk(Mrj + j) + Npb + b 76 TEO BANICA Thus the associated master Hadamard matrix is given by: i a Nk(Mrj +j)+Npb+b Hia,jb = (q w ) = νijqi(Npb+b)wa(Npb+b) = νijwabqi(Npb+b) ij ab i(Npb+b) Now since (FM ⊗ FN )ia,jb = ν w , we get H = FM ⊗Q FN with Qib = q , as claimed. Observe that Q itself is a “master matrix”, because the indices split. In view of the above examples, and of the lack of other known examples of master Hadamard matrices, he following conjecture was made in [2]: Conjecture 5.12 (Master Hadamard Conjecture). The master Hadamard matrices ap- pear as Dit¸˘adeformations of FN . There is a relation here with the notions of defect and isolation, that we would like to discuss now. First, we have the following defect computation: Theorem 5.13. The defect of a master Hadamard matrix is given by 1 d(H) = dim B ∈ M ( ) B¯ = BL, (BR) = (BR) ∀i, j R N C N i,ij j,ij 1 λj where Lij = f( ) and Ri,jk = f( ), f being the master function. λiλj λiλk Proof. The first order deformation equations are: X ¯ HikHjk(Aik − Ajk) = 0 k n j ¯ nk With Hij = λi we have HijHjk = (λi/λj) , and so the defect is given by: ( ) λ nk λ nk X i X i d(H) = dimR A ∈ MN (R) Aik = Ajk ∀i, j λj λj k k t 1 ¯ Now, pick A ∈ MN (C) and set B = AH , so that A = N BH. First, we have: ¯ ¯ A ∈ MN (R) ⇐⇒ BH = BH 1 ⇐⇒ B¯ = BHH¯ ∗ N On the other hand, the matrix on the right is given by: ¯ ∗ X ¯ ¯ X −nk (HH )ij = HikHjk = (λiλj) = Lij k k HADAMARD MATRICES 77 ¯ 1 Thus A ∈ MN (R) if and only the condition B = N BL in the statement is satisfied. 1 ¯ Regarding now the second condition on A, observe that with A = N BH we have: nk nk X λi 1 X λi Aik = Bis λj N λjλs k ks 1 X = B R N is s,ij s 1 = (BR) N i,ij Thus the second condition on A reads (BR)i,ij = (BR)j,ij, and we are done. In view of the above results, a conjecture would be that the only isolated master Ha- damard matrices are the Fourier matrices Fp, with p prime. See [22]. Let us discuss now yet another interesting construction of complex Hadamard matrices, due to McNulty and Weigert [69]. The matrices constructed there generalize the Tao matrix T6, and usually have the interesting feature of being isolated. The construction in [69] uses the theory of MUB, as developed in [26], [48], but we will follow here a more direct approach, using basic Gauss sums, from [22]. The starting observation from [69] is as follows: Theorem 5.14. Assuming that K ∈ MN (C) is Hadamard, so is the matrix 1 H = √ K (L∗R ) ia,jb Q ij i j ab √ √ provided that {L1,...,L√N } ⊂ QUQ and {R1,...,RN } ⊂ QUQ are such that each of √1 ∗ the matrices Q Li Rj ∈ QUQ, with i, j = 1,...,N, is Hadamard. Proof. The check of the unitarity is done as follows: 1 X < H ,H > = K (L∗R ) K¯ (L∗R ) ia kc Q ij i j ab kj k j cb jb X ¯ ∗ = KijKkj(Li Lk)ac j ∗ = Nδik(Li Lk)ac = NQδikδac The entries being in addition on the unit circle, we are done. As input for the above, we can use the following well-known Fourier construction: 78 TEO BANICA q−1 Proposition 5.15. For q ≥ 3 prime, the matrices {Fq,DFq,...,D Fq}, where 2 3 6 10 q −1 10 6 3 D = diag 1, 1, w, w , w , w , . . . , w 8 , . . . , w , w , w , w 2πi/q √1 ∗ with w = e , are such that q Ei Ej is complex Hadamard, for any i 6= j. Proof. With 0, 1, . . . , q − 1 as indices, the formula of the above matrix D is: 0+1+...+(c−1) c(c−1) Dc = w = w 2 √ √1 ∗ Since we have q Ei Ej ∈ qUq, we just need to check that these matrices have entries belonging to T, for any i 6= j. With k = j − i, these entries are given by: 1 1 1 X √ (E∗E ) = √ (F ∗DkF ) = √ wc(b−a)Dk q i j ab q q q ab q c c Now observe that with s = b − a, we have the following formula: 2 X cs k X cs−ds c(c−1) ·k− d(d−1) ·k 2 2 w Dc = w w c cd X (c−d) c+d−1 ·k+s = w ( 2 ) cd X e 2d+e−1 ·k+s = w ( 2 ) de ! X e(e−1) ·k+es X edk = w 2 w e d X e(e−1) ·k+es = w 2 · qδe0 e = q Thus the entries are on the unit circle, and we are done. We recall that the Legendre symbol is defined as follows: 0 if s = 0 s = 1 if ∃ α, s = α2 q −1 if 6∃ α, s = α2 Here, and in what follows, all the numbers are taken modulo q. We have: HADAMARD MATRICES 79 1 ∗ k c(c−1) √ 2 Proposition 5.16. The matrices Gk = q Fq D Fq, with D = diag(w ) being as above, and with k 6= 0 are circulant, their first row vectors V k being given by 2 i ( i −1) k k/2 q −1 ·k − k k V = δ w 8 · w 2 i q q where δq = 1 if q = 1(4) and δq = i if q = 3(4), and with all inverses being taken in Zq. Proof. This is a standard exercice on quadratic Gauss sums. First of all, the matrices Gk in the statement are indeed circulant, their first vectors being given by: k 1 X c(c−1) ·k+ic V = √ w 2 i q c Let us first compute the square of this quantity. We have: k 2 1 X c(c−1) + d(d−1) k+i(c+d) (V ) = w[ 2 2 ] i q cd The point now is that the sum S on the right, which has q2 terms, decomposes as follows, where x is a certain exponent, depending on q, i, k: ( (q − 1)(1 + w + ... + wq−1) + qwx if q = 1(4) S = (q + 1)(1 + w + ... + wq−1) − qwx if q = 3(4) We conclude that we have a formula as follows, where δq ∈ {1, i} is as in the statement, 2 2 2 so that δq ∈ {1, −1} is given by δq = 1 if q = 1(4) and δq = −1 if q = 3(4): k 2 2 x (Vi ) = δq w In order to compute now the exponent x, we must go back to the above calculation of the sum S. We succesively have: q2−1 – First of all, at k = 1, i = 0 we have x = 4 . q2−1 – By translation we obtain x = 4 − i(i − 1), at k = 1 and any i. k q2−1 i i – By replacing w → w we obtain x = 4 · k − k ( k − 1), at any k 6= 0 and any i. Summarizing, we have computed the square of the quantity that we are interested in, the formula being as follows, with δq being as in the statement: q2−1 i i k 2 2 4 ·k − k ( k −1) (Vi ) = δq · w · w By extracting now the square root, we obtain a formula as follows: 2 i ( i −1) q −1 k k k 8 ·k − 2 Vi = ±δq · w · w The computation of the missing sign is non-trivial, but by using the theory of quadratic Gauss sums, and more specifically a result of Gauss, computing precisely this kind of sign, k/2 we conclude that we have indeed a Legendre symbol, ± = q , as claimed. 80 TEO BANICA Let us combine now all the above results. We obtain the following statement: Theorem 5.17. Let q ≥ 3 be prime, consider two subsets S,T ⊂ {0, 1, . . . , q−1} satisfying |S| = |T | and S ∩ T = ∅, and write S = {s1, . . . , sN } and T = {t1, . . . , tN }. Then tj −si Hia,jb = KijVb−a where V is as above, is Hadamard, provided that K ∈ MN (C) is. Proof. This follows indeed by using the general construction in Theorem 5.14 above, with input coming from Proposition 5.15 and Proposition 5.16. As explained in [69], the above construction covers many interesting examples of Ha- damard matrices, known from [88], [89] to be isolated, such as the Tao matrix: 1 1 1 1 1 1 2 2 1 1 w w w w 2 2 1 w 1 w w w T6 = 2 2 1 w w 1 w w 1 w2 w2 w 1 w 1 w2 w w2 w 1 In general, in order to find isolated matrices, the idea from [69] is that of starting with an isolated matrix, and then use suitable sets S,T . The defect computations are, however, quite difficult. As a concrete statement, however, we have the following conjecture: Conjecture 5.18. The complex Hadamard matrix constructed in Theorem 5.17 is iso- lated, provided that: (1) K is an isolated Fourier matrix, of prime order. (2) S,T consist of consecutive odd numbers, and consecutive even numbers. This statement is supported by the isolation result for T6, and by several computer simulations in [69]. For further details on all this, we refer to [69], and to [22]. As a final topic now, we would like to discuss an extension of a part of the above results, to the case of the partial Hadamard matrices. The extension, done in [22], is quite straightforward, but there are however a number of subtleties appearing. First of all, we can talk about deformations of PHM, as follows: Definition 5.19. Let H ∈ XM,N be a partial complex Hadamard matrix. (1) A deformation of H is a smooth function f : T1 → (XM,N )H . Aij (2) The deformation is called “affine” if fij(q) = Hijq , with A ∈ MM×N (R). ai+bj M N (3) We call “trivial” the deformations fij(q) = Hijq , with a ∈ R , b ∈ R . √ We have XM,N = MM×N (T)∩ NUM,N , where UM,N ⊂ MM×N (C) is the set of matrices having all rows of norm 1, and pairwise orthogonal. This remark leads us to: HADAMARD MATRICES 81 Definition 5.20. Associated to a point H ∈ XM,N are: √ (1) The enveloping tangent space: TeH XM,N = TH MM×N (T) ∩ TH NUM,N . (2) The tangent cone TH XM,N : the set of tangent vectors to the deformations of H. ◦ (3) The affine tangent cone TH XM,N : same as above, using affine deformations only. × (4) The trivial tangent cone TH XM,N : as above, using trivial deformations only. ◦ Observe that TeH XM,N ,TH XM,N are real vector spaces, and that TH XM,N ,TH XM,N are two-sided cones, λ ∈ R,A ∈ T =⇒ λA ∈ T . Also, we have inclusions as follows: × ◦ TH XM,N ⊂ TH XM,N ⊂ TH XM,N ⊂ TeH XM,N Since TeH XM,N is a real vector space, of particular interest is the computation of its dimension d(H) = dim(TeH XM,N ), called defect of H. We have: √ Theorem 5.21. Let H ∈ XM,N , and pick K ∈ NUN extending H. P ¯ (1) TeH XM,N '{A ∈ MM×N (R)| k HikHjk(Aik − Ajk) = 0, ∀i, j}. ∗ ¯ (2) TeH XM,N '{E = (XY ) ∈ MM×N (C)|X = X , (EK)ijHij ∈ R, ∀i, j}. P ¯ ¯ The correspondence A → E is given by Eij = k HikKjkAik, Aij = (EK)ijHij. Proof. The proofs here go as in the square case, as follows: (1) In the square case this was done in the proof of Theorem 4.4 above, and the extension of the computations there to the rectangular case is straightforward. ∗ (2) Let us set indeed Rij = AijHij and E = RK . The correspondence A → R → E is then bijective, and we have the following formula: X ¯ Eij = HikKjkAik k ¯ With these changes, the system of equations in (1) becomes Eij = Eji for any i, j with j ≤ M. But this shows that we must have E = (XY ) with X = X∗, and the condition ¯ Aij ∈ R corresponds to the condition (EK)ijHij ∈ R, as claimed. As an illustration, in the real case we obtain the following result: Theorem 5.22. For an Hadamard matrix H ∈ MM×N (±1) we have symm TeH XM,N ' MM (R) ⊕ MM×(N−M)(R) and so the defect is given by d(H) = N(N + 1)/2 + M(N − M) independently of the precise value of H. 82 TEO BANICA √ Proof. We use Theorem 5.21 (2). Since H is now real we can pick K ∈ NUN extending it to be real too, and with nonzero entries, so the last condition appearing there, namely ¯ (EK)ijHij ∈ R, simply tells us that E must be real. Thus we have: ∗ TeH XM,N '{E = (XY ) ∈ MM×N (R)|X = X } But this is the formula in the statement, and we are done. A matrix H ∈ XM,N cannot be isolated, simply because the space of its Hadamard MN equivalents provides a copy T ⊂ XM,N , passing through H. However, if we restrict the attention to the matrices which are dephased, the notion of isolation makes sense: Proposition 5.23. Let d(H) = dim(TeH XM,N ). (1) This number, called undephased defect of H, satisfies d(H) ≥ M + N − 1. (2) If d(H) = M+N−1 then H is isolated inside the dephased quotient XM,N → ZM,N . Proof. Once again, the known results in the square case extend: × (1) We have indeed dim(TH XM,N ) = M + N − 1, and since the tangent vectors to these trivial deformations belong to TeH XM,N , this gives the result. × (2) Since d(H) = M + N − 1, the inclusions TH XM,N ⊂ TH XM,N ⊂ TeH XM,N must be × equalities, and from TH XM,N = TH XM,N we obtain the result. Finally, still at the theoretical level, we have the following conjecture: Conjecture 5.24. An isolated matrix H ∈ ZM,N must have minimal defect, namely d(H) = M + N − 1. In other words, the conjecture is that if H ∈ ZM,N has only trivial first order deforma- tions, then it has only trivial deformations at any order, including at ∞. In the square matrix case this statement comes with solid evidence, all known examples of complex Hadamard matrices H ∈ XN having non-minimal defect being known to admit one-parameter deformations. For more on this subject, see [88], [89]. Let us discuss now some examples of isolated partial Hadamard matrices, and provide some evidence for Conjecture 5.24. We are interested in the following matrices: Definition 5.25. The truncated Fourier matrix FS,G, with G being a finite abelian group, and with S ⊂ G being a subset, is constructed as follows: ij 2πi/N (1) Given N ∈ N, we set FN = (w )ij, where w = e . (2) Assuming G = ZN1 × ... × ZNs , we set FG = FN1 ⊗ ... ⊗ FNs . (3) We let FS,G be the submatrix of FG having S ⊂ G as row index set. Observe that FN is the Fourier matrix of the cyclic group ZN . More generally, FG is the Fourier matrix of the finite abelian group G. Observe also that FG,G = FG. We can compute the defect of FS,G by using Theorem 5.21, and we obtain: HADAMARD MATRICES 83 Theorem 5.26. For a truncated Fourier matrix F = FS,G we have the formula n o t ¯ TeF XM,N = A ∈ MM×N (R) P = AF satisfies Pij = Pi+j,j = Pi,−j, ∀i, j where M = |S|,N = |G|, and with all the indices being regarded as group elements. Proof. We use Theorem 5.21 (1). The defect equations there are as follows: X ¯ FikFjk(Aik − Ajk) = 0 k ¯ t Since for F = FS,G we have FikFjk = (F )k,i−j, we obtain: n o t t TeF XM,N = A ∈ MM×N (R) (AF )i,i−j = (AF )j,i−j, ∀i, j Now observe that for an arbitrary matrix P ∈ MM (C), we have: Pi,i−j = Pj,i−j, ∀i, j ⇐⇒ Pi+j,i = Pji, ∀i, j ⇐⇒ Pi+j,j = Pij, ∀i, j We therefore conclude that we have the following equality: n o t TeF XM,N = A ∈ MM×N (R) P = AF satisfies Pij = Pi+j,j, ∀i, j t Now observe that with A ∈ MM×N (R) and P = AF ∈ MM (C) as above, we have: ¯ X ∗ Pij = Aik(F )kj k X t = Aik(F )k,−j k = Pi,−j Thus, we obtain the formula in the statement, and we are done. Let us try to find some explicit examples of isolated matrices, of truncated Fourier type. For this purpose, we can use the following improved version of Theorem 5.26: Theorem 5.27. The defect of F = FS,G is the number dim(K) + dim(I), where n o t K = A ∈ MM×N (R) AF = 0 n o t I = P ∈ LM ∃A ∈ MM×N (R),P = AF where LM is the following linear space ¯ LM = P ∈ MM (C) Pij = Pi+j,j = Pi,−j, ∀i, j with all the indices belonging by definition to the group G. 84 TEO BANICA Proof. We use the general formula in Theorem 5.26. With the notations there, and with the linear space LM being as above, we have a linear map as follows: t Φ: TeF XM,N → LM , Φ(A) = AF By using this map, we obtain the following equality: dim(TeF XM,N ) = dim(ker Φ) + dim(Im Φ) Now since the spaces on the right are precisely those in the statement, ker Φ = K and Im Φ = I, by applying Theorem 5.26 we obtain the result. In order to look now for isolated matrices, the first remark is that since a deformation of FG will produce a deformation of FS,G too, we must restrict the attention to the case where G = Zp, with p prime. And here, we have the following conjecture: Conjecture 5.28. There exists a constant ε > 0 such that FS,p is isolated, for any p prime, once S ⊂ Zp satisfies |S| ≥ (1 − ε)p. In principle this conjecture can be approached by using the formula in Theorem 5.27, and we have for instance evidence towards the fact that Fp−1,p should be always isolated, that Fp−2,p should be isolated too, provided that p is big enough, and so on. However, finding a number ε > 0 as above looks like a quite difficult question. See [22]. HADAMARD MATRICES 85 6. Circulant matrices We discuss in this section yet another type of special complex Hadamard matrices, namely the circulant ones. There has been a lot of work here, starting with the Circulant Hadamard Conjecture (CHC) in the real case, and with many results in the complex case as well. We will present here the main techniques in dealing with such matrices. It is convenient to introduce the circulant matrices as follows: Definition 6.1. A complex matrix H ∈ MN (C) is called circulant when we have Hij = γj−i for some γ ∈ CN , with the matrix indices i, j ∈ {0, 1,...,N − 1} taken modulo N. Here the index convention is quite standard, as for the Fourier matrices FN , and with this coming from Fourier analysis considerations, that we will get into later on. Here is a basic, and very fundamental example of a circulant Hadamard matrix, which in addition has real entries, and is symmetric: −1 1 1 1 1 −1 1 1 K4 = 1 1 −1 1 1 1 1 −1 According to the CHC, explained in section 1, this matrix is, up to equivalence, the only circulant Hadamard matrix H ∈ MN (±1), regardless of the value of N ∈ N. Our first purpose will be that of showing that the CHC dissapears in the complex case, where we have examples at any N ∈ N. As a first result here, we have: Proposition 6.2. The following are circulant and symmetric Hadamard matrices, −1 ν 1 ν w 1 1 i 1 ν −1 ν 1 F 0 = ,F 0 = 1 w 1 ,F 00 = 2 1 i 3 4 1 ν −1 ν 1 1 w ν 1 ν −1 2πi/3 πi/4 where w = e , ν = e , equivalent to the Fourier matrices F2,F3,F4. Proof. The orthogonality between rows being clear, we have here complex Hadamard 0 matrices. The fact that we have an equivalence F2 ∼ F2 follows from: 1 1 i i i 1 ∼ ∼ 1 −1 1 −1 1 i 0 At N = 3 now, the equivalence F3 ∼ F3 can be constructed as follows: 1 1 1 1 1 w w 1 1 1 w w2 ∼ 1 w 1 ∼ 1 w 1 1 w2 w w 1 1 1 1 w 86 TEO BANICA 00 As for the case N = 4, here the equivalence F4 ∼ F4 can be constructed as follows, 2πki/s where we use the logarithmic notation [k]s = e , with respect to s = 8: 0 0 0 0 0 1 4 1 4 1 0 1 0 2 4 6 1 4 1 0 1 4 1 0 ∼ ∼ 0 4 0 4 4 1 0 1 0 1 4 1 0 6 4 2 1 0 1 4 1 0 1 4 8 8 8 00 0 We will explain later the reasons for denoting this matrix F4 , instead of F4. Getting back now to the real circulant matrix K4, this is equivalent to the Fourier matrix FG = F2 ⊗ F2 of the Klein group G = Z2 × Z2, as shown by: −1 1 1 1 1 1 1 −1 1 1 1 1 1 −1 1 1 1 −1 1 1 1 −1 1 −1 ∼ ∼ 1 1 −1 1 1 1 −1 1 1 1 −1 −1 1 1 1 −1 −1 1 1 1 1 −1 −1 1 In fact, we have the following construction of circulant and symmetric Hadamard ma- trices at N = 4, which involves an extra parameter q ∈ T: Proposition 6.3. The following circulant and symmetric matrix is Hadamard, −1 q 1 q q q −1 q 1 K = 4 1 q −1 q q 1 q −1 πi/4 00 for any q ∈ T. At q = 1, e recover respectively the matrices K4,F4 . Proof. The rows of the above matrix are pairwise orthogonal for any q ∈ C, and so at q ∈ T we obtain a complex Hadamard matrix. The last assertion is clear. As a first conclusion, coming from the above considerations, we have: Theorem 6.4. The complex Hadamard matrices of order N = 2, 3, 4, 5, namely p F2,F3,F4 ,F5 can be put, up to equivalence, in circulant and symmetric form. Proof. As explained in section 2 above, according to the result of Haagerup [51], the Hadamard matrices at N = 2, 3, 4, 5 are, up to equivalence, those in the statement. At N = 2, 3 the problem is solved by Proposition 6.2 above. HADAMARD MATRICES 87 q s −2 At N = 4 now, our claim is that we have K4 ∼ F4 , with s = q . Indeed, by multiplying q the rows of K4 , and then the columns, by suitable scalars, we have: −1 q 1 q 1 −q −1 −q 1 1 1 1 q q −1 q 1 1 −q¯ 1q ¯ 1 s −1 −s K = ∼ ∼ 4 1 q −1 q 1 q −1 q 1 −1 1 −1 q 1 q −1 1q ¯ 1 −q¯ 1 −s −1 s s On the other hand, by permuting the second and third rows of F4 , we obtain: 1 1 1 1 1 1 1 1 s 1 −1 1 −1 1 s −1 −s F = ∼ 4 1 s −1 −s 1 −1 1 −1 1 −s −1 s 1 −s −1 s Thus these matrices are equivalent, and the result follows from Proposition 6.3. At N = 5, the matrix that we are looking for is as follows, with w = e2πi/5: w2 1 w4 w4 1 1 w2 1 w4 w4 0 4 2 4 F5 = w 1 w 1 w 4 4 2 w w 1 w 1 1 w4 w4 1 w2 It is indeed clear that this matrix is circulant, symmetric, and complex Hadamard, and 0 the fact that we have F5 ∼ F5 follows either directly, or by using [51]. Let us prove now that any Fourier matrix FN can be put in circulant and symmetric form. We use Bj¨orck’s cyclic root formalism [30], which is as follows: Theorem 6.5. Assume that H ∈ MN (T) is circulant, Hij = γj−i. Then H is Hadamard if and only if the vector (z0, z1, . . . , zN−1) given by zi = γi/γi−1 satisfies: z0 + z1 + ... + zN−1 = 0 z0z1 + z1z2 + ... + zN−1z0 = 0 ... z0z1 . . . zN−2 + ... + zN−1z0 . . . zN−3 = 0 z0z1 . . . zN−1 = 1 If so is the case, we say that z = (z0, . . . , zN−1) is a cyclic N-root. Proof. This follows from a direct computation, the idea being that, with Hij = γj−i as above, the orthogonality conditions between the rows are best written in terms of the variables zi = γi/γi−1, and correspond to the equations in the statement. See [30]. 88 TEO BANICA Observe that, up to a global multiplication by a scalar w ∈ T, the first row vector γ = (γ0, . . . , γN−1) of the matrix H ∈ MN (T) constructed above is as follows: γ = (z0, z0z1, z0z1z2, ...... , z0z1 . . . zN−1) Now back to the Fourier matrices, we have the following result: Theorem 6.6. Given N ∈ N, set ν = eπi/N and q = νN−1, w = ν2. Then (q, qw, qw2, . . . , qwN−1) 0 is a cyclic N-root, and the corresponding complex Hadamard matrix FN is circulant and symmetric, and equivalent to the Fourier matrix FN . Proof. Given q, w ∈ T, let us find out when (q, qw, qw2, . . . , qwN−1) is a cyclic root: (1) In order for the = 0 equations in Theorem 6.5 to be satisfied, the value of q is irrelevant, and w must be a primitive N-root of unity. (2) As for the = 1 equation in Theorem 6.5, this states in our case that we must have N N(N−1) N N−1 q w 2 = 1, and so that we must have q = (−1) . We conclude that with the values of q, w ∈ T in the statement, we have indeed a cyclic N-root. Now construct Hij = γj−i as in Theorem 6.5. We have: k+1 k(k+1) −k+1 k(k−1) γk = γ−k, ∀k ⇐⇒ q w 2 = q w 2 , ∀k ⇐⇒ q2kwk = 1, ∀k ⇐⇒ q2 = w−1 But this latter condition holds indeed, because we have q2 = ν2N−2 = ν−2 = w−1. We conclude that our circulant matrix H is symmetric as well, as claimed. It remains to construct an equivalence H ∼ FN . In order to do this, observe that, due to our conventions q = νN−1, w = ν2, the first row vector of H is given by: k+1 k(k+1) γk = q w 2 = ν(N−1)(k+1)νk(k+1) = ν(N+k−1)(k+1) Thus, the entries of H are given by the following formula: H−i,j = H0,i+j = ν(N+i+j−1)(i+j+1) = νi2+j2+2ij+Ni+Nj+N−1 = νN−1 · νi2+Ni · νj2+Nj · ν2ij With this formula in hand, we can now finish. Indeed, the matrix H = (Hij) is 0 0 equivalent to the matrix H = (H−i,j). Now regarding H , observe that in the above formula, the factors νN−1, νi2+Ni, νj2+Nj correspond respectively to a global multiplication HADAMARD MATRICES 89 by a scalar, and to row and column multiplications by scalars. Thus H0 is equivalent to the matrix H00 obtained by deleting these factors. 00 2ij πi/N But this latter matrix, given by Hij = ν with ν = e , is precisely the Fourier matrix FN , and we are done. As an illustration, let us work out the cases N = 2, 3, 4, 5. We have here: 0 Proposition 6.7. The matrices FN are as follows: 0 0 (1) At N = 2, 3 we obtain the old matrices F2,F3. (2) At N = 4 we obtain the following matrix, with ν = eπi/4: ν3 1 ν7 1 3 7 0 1 ν 1 ν F = 7 3 4 ν 1 ν 1 1 ν7 1 ν3 0 (3) At N = 5 we obtain the old matrix F5. Proof. With notations from Theorem 6.6, the proof goes as follows: (1) At N = 2 we have ν = i, q = i, w = −1, so the cyclic root is (i, −i), the first row 0 πi/3 vector is (i, 1), and we obtain indeed the old matrix F2. At N = 3 we have ν = e and q = w = ν2 = e2πi/3, the cyclic root is (w, w2, 1), the first row vector is (w, 1, 1), and we 0 obtain indeed the old matrix F3. (2) At N = 4 we have ν = eπi/4 and q = ν3, w = ν2, the cyclic root is (ν3, ν5, ν7, ν), the first row vector is (ν3, 1, ν7, 1), and we obtain the matrix in the statement. (3) At N = 5 we have ν = eπi/5 and q = ν4 = w2, with w = ν2 = e2πi/5, the cyclic root is therefore (w2, w3, w4, 1, w), the first row vector is (w2, 1, w4, w4, 1), and we obtain 0 in this way the old matrix F5, as claimed. 0 00 Regarding the above matrix F4, observe that this is equivalent to the matrix F4 from 0 00 Proposition 6.2, with the equivalence F4 ∼ F4 being obtained by multiplying everything by ν = eπi/4. While both these matrices are circulant and symmetric, and of course equiv- 0 alent to F4, one of them, namely F4, is “better” than the other, because the corresponding 0 00 cyclic root comes from a progression. This is the reason for our notations F4,F4 . Let us discuss now the case of the generalized Fourier matrices FG. In this context, the assumption of being circulant is somewhat unnatural, because this comes from a ZN symmetry, and the underlying group is no longer ZN . It is possible to fix this issue by talking about G-patterned Hadamard matrices, with G being no longer cyclic, but for our purposes here, best is to formulate the result in a weaker form, as follows: Theorem 6.8. The generalized Fourier matrices FG, associated to the finite abelian groups G, can be put in symmetric and bistochastic form. 90 TEO BANICA Proof. We know from Theorem 6.6 that any usual Fourier matrix FN can be put in circulant and symmetric form. Since circulant implies bistochastic, in the sense that the sums on all rows and all columns must be equal, the result holds for FN . In general now, if we decompose G = ZN1 × ... × ZNk , we have: FG = FN1 ⊗ ... ⊗ FNk Now since the property of being circulant is stable under taking tensor products, and so is the property of being bistochastic, we therefore obtain the result. We have as well the following alternative generalization of Theorem 6.6, coming from Backelin’s work in [3], and remaining in the circulant and symmetric setting: Theorem 6.9. Let M|N, and set w = e2πi/N . We have a cyclic root as follows, N−1 N−1 (q1, . . . , qM , q1w, . . . , qM w, ...... , q1w , . . . , qM w ) | {z } | {z } | {z } M M M N M(N−1) provided that q1, . . . , qM ∈ T satisfy (q1 . . . qM ) = (−1) . Moreover, assuming q1q2 = 1 , q3qM = q4qM−1 = ... = w N M(N−1) which imply (q1 . . . qM ) = (−1) , the Hadamard matrix is symmetric. Proof. Let us first check the = 0 equations for a cyclic root. Given arbitrary numbers q1, . . . , qM ∈ T, if we denote by (zi) the vector in the statement, we have: X q1 . . . qK + q2 . . . qK+1 + ...... + qM−K+1 . . . qM zi+1 . . . zi+K = K−1 +qM−K+2 . . . qM q1w + ...... + qM q1 . . . qK−1w i ×(1 + wK + w2K + ... + w(N−1)K ) Now since the sum on the right vanishes, the = 0 conditions are satisfied. Regarding now the = 1 condition, the total product of the numbers zi is given by: Y N 2 N−1 M N MN(N−1) zi = (q1 . . . qM ) (1 · w · w . . . w ) = (q1 . . . qM ) w 2 i By using w = e2πi/N we obtain that the coefficient on the right is: MN(N−1) 2πi · MN(N−1) πiM(N−1) M(N−1) w 2 = e N 2 = e = (−1) N M(N−1) Thus, if (q1 . . . qM ) = (−1) , we obtain a cyclic root, as stated. See [3], [49]. The corresponding first row vector can be written as follows: wM−1 w2 w V = q1, q1q2, . . . , q1 . . . qM ,...... , ,..., , , 1 | {z } q2 . . . qM qM−1qM qM M | {z } M HADAMARD MATRICES 91 Thus, the corresponding circulant complex Hadamard matrix is as follows: q1 q1q2 q1q2q3 q1q2q3q4 q1q2q3q4q5 ... 1 q1 q1q2 q1q2q3 q1q2q3q4 ... w 1 q1 q1q2 q1q2q3 ... qM H = w2 w 1 q1 q1q2 ... qM−1qM qM w3 w2 w 1 q1 ... qM−2qM−1qM qM−1qM qM ...... We are therefore led to the symmetry conditions in the statement, and we are done. Observe that Theorem 6.9 generalizes both Proposition 6.3, and the construction in Theorem 6.6. Thus, we have here a full generalization of Theorem 6.4. Of course, things do not stop here, and the problem of unifying Theorem 6.8 and Theorem 6.9 remains. As a conclusion to what we have so far, there is definitely no analogue of the CHC in the general complex setting, due to the fact that FN can be put in circulant form, and this latter result gives rise to a number of interesting generalizations and questions. However, still in relation with the CHC, but at a more technical level, the problem of investigating the existence of the circulant Butson matrices appears. The first result in this direction, due to Turyn [93], is as follows: Proposition 6.10. The size of a circulant Hadamard matrix H ∈ MN (±1) must be of the form N = 4n2, with n ∈ N. Proof. Let a, b ∈ N with a + b = N be the number of 1, −1 entries in the first row of H. If we denote by H0,...,HN−1 the rows of H, then by summing over columns we get: N−1 X < H0,Hi > = a(a − b) + b(b − a) i=0 = (a − b)2 On the other hand, the quantity on the left is < H0,H0 >= N. Thus N is a square, 2 and together with the fact that N ∈ 2N, this gives N = 4n , with n ∈ N. Also found by Turyn in [93] is the fact that the above number n ∈ N must be odd, and not a prime power. In the general Butson matrix setting now, we have: 2πi/l Proposition 6.11. Assume that H ∈ HN (l) is circulant, let w = e . If a0, . . . , al−1 ∈ P l−1 N with ai = N are the number of 1, w, . . . , w entries in the first row of H, then: X k w aiai+k = N ik P This condition, with ai = N, will be called “Turyn obstruction” on (N, l). 92 TEO BANICA Proof. Indeed, by summing over the columns of H, we obtain: X X i j < H0,Hi > = < w , w > aiaj i ij X i−j = w aiaj ij Now since the left term is < H0,H0 >= N, this gives the result. We can deduce from this a number of concrete obstructions, as follows: P 2 Theorem 6.12. When l is prime, the Turyn obstruction is i(ai − ai+k) = 2N for any k 6= 0. Also, for small values of l, the Turyn obstruction is as follows: 2 (1) At l = 2 the condition is (a0 − a1) = N. 2 2 2 (2) At l = 3 the condition is (a0 − a1) + (a1 − a2) + (a2 − a3) = 2N. 2 2 (3) At l = 4 the condition is (a0 − a2) + (a1 − a3) = N. P 2 P 2 (4) At l = 5 the condition is i(ai − ai+1) = i(ai − ai+2) = 2N. Proof. We use the fact, from Proposition 3.3 above, that when l is prime, the vanishing sums of l-roots of unity are exactly the sums of the following type, with c ∈ N: S = c + cw + ... + cwl−1 Thus the Turyn obstruction is equivalent to the following equations, one for each k 6= 0: X 2 X ai − aiai+k = N i i Now by forming squares, this gives the equations in the statement. Regarding now the l = 2, 3, 4, 5 assertions, these follow from the first assertion when l is prime, l = 2, 3, 5. Also, at l = 4 we have w = i, so the Turyn obstruction reads: 2 2 2 2 X X (a0 + a1 + a2 + a3) + i aiai+1 − 2(a0a2 + a1a3) − i aiai+1 = N Thus the imaginary terms cancel, and we obtain the formula in the statement. The above results are of course just some basic, elementary observations on the subject, and the massive amount of work on the CHC has a number of interesting Butson matrix extensions. For some more advanced theory on all this, we refer to [18], [39]. Let us go back now to the pure complex case, and discuss Fourier analytic aspects. From a traditional linear algebra viewpoint, the circulant matrices are best understood as being the matrices which are Fourier-diagonal,√ and we will exploit this here. ij 2πi/N Let us fix N ∈ N, and denote√ by F = (w )/ N with w = e the rescaled Fourier matrix. Observe that FN = NF is the usual Fourier Hadamard matrix. N Given a vector q ∈ C , we denote by Q ∈ MN (C) the diagonal matrix having q as vector of diagonal entries. That is, Qii = qi, and Qij = 0 for i 6= j. With these conventions, the above-mentioned linear algebra result is as follows: HADAMARD MATRICES 93 Theorem 6.13. For a complex matrix H ∈ MN (C), the following are equivalent: N (1) H is circulant, Hij = ξj−i for some ξ ∈ C . (2) H is Fourier-diagonal, H = F QF ∗ with Q diagonal. √ In addition, the first row vector of F QF ∗ is given by ξ = F q/ N. ∗ Proof. If Hij = ξj−i is circulant then Q = F HF is diagonal, given by: 1 X X Q = wjl−ikξ = δ wjrξ ij N l−k ij r kl r Also, if Q = diag(q) is diagonal then H = F QF ∗ is circulant, given by: X 1 X H = F Q F¯ = w(i−j)kq ij ik kk jk N k k k √ Observe that this latter formula proves as well the last assertion, ξ = F q/ N. In relation now with the orthogonal and unitary matrices, we have: Proposition 6.14. The various sets of circulant matrices are as follows: circ ∗ N (1) MN (C) = {F QF |q ∈ C }. circ ∗ N (2) UN = {F QF |q ∈ T }. (3) Ocirc = {F QF ∗|q ∈ N , q¯ = q , ∀i}. N T i −i √ In addition, the first row vector of F QF ∗ is given by ξ = F q/ N. Proof. All this follows from Theorem 6.13, as follows: (1) This assertion, along with the last one, is Theorem 6.13 itself. (2) This is clear from (1), because the eigenvalues must be on the unit circle T. N (3) Observe first that for q ∈ C we have F q = F q˜, withq ˜i =q ¯−i, and so ξ = F q is real if and only ifq ¯i = q−i for any i. Together with (2), this gives the result. Observe that in (3), the equations for the parameter space are q0 =q ¯0,q ¯1 = qn−1, q¯2 = qn−2, and so on until [N/2] + 1. Thus, with the convention Z∞ = T we have: ( × (N−1)/2 (N odd) Ocirc ' Z2 Z∞ N 2 (N−2)/2 Z2 × Z∞ (N even) In terms of circulant Hadamard matrices, we have the following statement: Theorem 6.15. The sets of complex and real circulant Hadamard matrices are: √ circ ∗ N XN = { NF QF |q ∈ T } ∩ MN (T) √ circ ∗ N YN = { NF QF |q ∈ T , q¯i = q−i} ∩ MN (±1) In addition, the sets of q parameters are invariant under cyclic permutations, and also under mutiplying by numbers in T, respectively under multiplying by −1. 94 TEO BANICA Proof. All the assertions are indeed clear from Proposition 6.14 above. The above statement is of course something quite theoretical in the real case, where circ the CHC states that we should have YN = ∅, at any N 6= 4. However, in the complex case all this is useful, and complementary to Bj¨orck’s cyclic root formalism. Let us discuss now a number of geometric and analytic aspects. First, we have the following deep counting result, due to Haagerup [52]: Theorem 6.16. When N is prime, the number of circulant N × N complex Hadamard 2N−2 matrices, counted with certain multiplicities, is exactly N−1 . Proof. This is something advanced, using a variety of techiques from Fourier analysis, number theory, complex analysis and algebraic geometry. The idea in [52] is that, when N is prime, Bj¨orck’s cyclic root formalism can be further manipulated, by using Fourier transforms, and we are eventually led to a simpler system of equations. This simplified system can be shown then to have a finite number of solutions, the key ingredient here being a well-known theorem of Chebotarev, which states that when N is prime, all the minors of the Fourier matrix FN are nonzero. With this finiteness result in hand, the precise count can be done as well, by using various techniques from classical algebraic geometry. See [52]. When N is not prime, the situation is considerably more complicated, with some values leading to finitely many solutions, and with other values leading to an infinite number of solutions, and with many other new phenomena appearing. See [30], [31], [32], [52]. Our belief is that useful here would be an adaptation of the notion of defect, to the circulant case, in the context of the manifolds from Proposition 6.14 and Theorem 6.15. There are some preliminary differential geometry computations to be done here. We would like to√ discuss now an alternative take on these questions, based on the estimate ||U||1 ≤ N N from Theorem 1.18. This shows that the real Hadamard matrices are the rescaled versions of the maximizers of the 1-norm on ON , and the same proof shows that the complex Hadamard matrices are the rescaled versions of the maximizers of the 1-norm on UN . Following [18], we will apply this philosophy to the circulant case. We will need in fact more general p-norms as well, so let us start with the following result, in the complex case, which is the most general one on the subject: Proposition 6.17. If ψ : [0, ∞) → R is strictly concave/convex, the quantity X 2 F (U) = ψ(|Uij| ) ij over UN is maximized/minimized precisely by the rescaled Hadamard matrices. HADAMARD MATRICES 95 Proof. We recall that Jensen’s inequality states that for ψ convex we have: x + ... + x ψ(x ) + ... + ψ(x ) ψ 1 n ≤ 1 n n n For ψ concave the reverse inequality holds. Also, the equality case holds either when ψ is linear, or when the numbers x1, . . . , xn are all equal. 2 2 In our case, with n = N and with {x1, . . . , xn} = {|Uij| |i, j = 1,...,N}, we obtain that for any convex function ψ, the following holds: 1 F (U) ψ ≤ N N 2 Thus we have F (U) ≥ N 2ψ(1/N), and by assuming as in the statement that ψ is 2 strictly convex,√ the equality case holds precisely when the numbers |Uij| are all equal, so when H = NU is Hadamard. The proof for concave functions is similar. Of particular interest for our considerations are the power functions ψ(x) = xp/2, which are concave at p ∈ [1, 2), and convex at p ∈ (2, ∞). These lead to: √ Theorem 6.18. The rescaled versions U = H/ N of the complex Hadamard matrices H ∈ MN (C) can be characterized as being: (1) The maximizers of the p-norm on UN , at any p ∈ [1, 2). (2) The minimizers of the p-norm on UN , at any p ∈ (2, ∞]. Proof. Consider indeed the p-norm on UN , which at p ∈ [1, ∞) is given by: !1/p X p ||U||p = |Uij| ij By the above discussion, involving the functions ψ(x) = xp/2, Proposition 6.17 applies and gives the results at p ∈ [1, ∞), the precise estimates being as follows: ≤ N 2/p−1/2 if p < 2 1/2 ||U||p = = N if p = 2 ≥ N 2/p−1/2 if p > 2 As for the case p = ∞, this follows with p → ∞, or directly via Cauchy-Schwarz. As explained in [18], the most adapted exponent for the circulant case is p = 4, due to a number of simplifications which appear in the Fourier manipulations. So, before discussing this, let us record the p = 4 particular case of Theorem 6.18: Proposition 6.19. Given a matrix U ∈ UN we have ||U||4 ≥ 1 √ with equality precisely when H = U/ N is Hadamard. 96 TEO BANICA Proof. This follows from Theorem 6.18, or directly from Cauchy-Schwarz, as follows: !2 X 1 X ||U||4 = |U |4 ≥ |U |2 = 1 4 ij N 2 ij ij ij √ Thus we have ||U||4 ≥ 1, with equality if and only if H = NU is Hadamard. In the circulant case now, and in Fourier formulation, the estimate is as follows: N Theorem 6.20. Given a vector q ∈ T , written q = (q0, . . . , qN−1) consider the following quantity, with all the indices being taken modulo N: X qiqk Φ = qjql i+k=j+l √ Then Φ is real, and we have Φ ≥ N 2, with equality if and only if Nq is the eigenvalue vector of a circulant Hadamard matrix H ∈ MN (C). Proof. By conjugating the formula of Φ we see that this quantity is indeed real. In fact, Φ appears by definition as a sum of N 3 terms, consisting of N(2N − 1) values of 1 and of N(N − 1)2 other complex numbers of modulus 1, coming in pairs (a, a¯). Regarding now the second assertion, by using the√ various identifications in Theorem 6.13 and Proposition 6.14, and the formula ξ = F q/ N there, we have: 4 X 4 ||U||4 = N |ξs| s 1 X X = | wsiq |4 N 3 i s i 1 X X X X X = wsiq w−sjq¯ wskq w−slq¯ N 3 i j k l s i j k l 1 X X (i−j+k−l)s qiqk = 3 w N qjql s ijkl 1 X qiqk = 2 N qjql i+k=j+l Thus Proposition 6.19 gives the following estimate: 2 4 2 Φ = N ||U||4 ≥ N Moreover, we have equality precisely in the Hadamard matrix case, as claimed. We have the following more direct explanation of the above result: HADAMARD MATRICES 97 Proposition 6.21. With the above notations, we have the formula 2 X 2 2 2 Φ = N + (|νi| − |νj| ) i6=j where ν = (ν0, . . . , νN−1) is the vector given by ν = F q. Proof. This follows by replacing in the above proof the Cauchy-Schwarz estimate by the corresponding sum of squares. More precisely, we know from the above proof that: 3 X 4 Φ = N |ξi| i P 2 On the other hand Uij = ξj−i being unitary, we have i |ξi| = 1, and so: X 4 X 2 2 1 = |ξi| + |ξi| · |ξj| i i6=j ! X 4 X 4 X 2 2 = N |ξi| − (N − 1) |ξi| − |ξi| · |ξj| i i i6=j 1 X = Φ − (|ξ |2 − |ξ |2)2 N 2 i j i6=j 2 Now by multiplying by N , this gives the formula in the statement. All this is quite interesting. As an application, in the real Hadamard matrix case, we have the following analytic reformulation of the CHC, from [18]: N Theorem 6.22. For q ∈ T satisfying q¯i = q−i, the following quantity is real, X Φ = qiqjqkql i+j+k+l=0 and satisfies Φ ≥ N 2. The CHC states that we cannot have equality at N > 4. Proof. This follows indeed from Theorem 6.20, via the identifications from Theorem 6.15, N the parameter space in the real case being {q ∈ T |q¯i = q−i}. This is certainly quite nice, and the analytic problem might look quite elementary. However, this is not the case. In fact, we already know from section 1 that the CHC is equivalent to Ryser’s conjecture, which looks elementary as well, and is not. Following [18], let us further discuss all this. We first have: Theorem 6.23. Write Φ = Φ0 +...+ΦN−1, with each Φi being given by the same formula as Φ, namely Φ = P qiqk , but keeping the index i fixed. Then: i+k=j+l qj qk (1) The critical points of Φ are those where Φi ∈ R, for any i. (2) In the Hadamard case we have Φi = N, for any i. 98 TEO BANICA Proof. This follows by doing some elementary computations, as follows: (1) The first observation is that the non-constant terms in the definition of Φ involving ¯ the variable qi are the terms of the sum Ki + Ki, where: 2 X qi X X qiqk Ki = + 2 qjql qjql 2i=j+l k6=i i+k=j+l iαi Thus if we fix i and we write qi = e , we obtain: ! ∂Φ X X qiqk = 4Re i · ∂αi qjql k i+k=j+l ! X qiqk = 4Im qjql i+k=j+l = 4Im(Φi) Now since the derivative must vanish for any i, this gives the result. (2) We first perform the end of the Fourier computation in the proof of Theorem 6.20 above backwards, by keeping the index i fixed. We obtain: X qiqk Φi = qjql i+k=j+l 1 X X qiqk = w(i−j+k−l)s N qjql s ijkl 1 X X X X = wsiq w−sjq¯ wskq w−slq¯ N i j k l s j k l 2 X si ¯ ¯ = N w qiξsξsξs s √ Here we have used the formula√ξ = F q/ N. Now by assuming that we are in the Hadamard case, we have |ξs| = 1/ N for any s, and so we obtain: √ X si ¯ ∗ Φi = N w qiξs = N Nqi(F ξ)i = Nqiq¯i = N s Thus, we have obtained the conclusion in the statement. Let us discuss now a probabilistic approach to all this. Given a compact manifold X endowed with a probability measure, and a bounded function Θ : X → [0, ∞), the maximum of this function can be recaptured via following well-known formula: Z 1/p max Θ = lim Θ(x)p dx p→∞ X HADAMARD MATRICES 99 In our case, we are rather interested in computing a minimum, and the result is: Proposition 6.24. We have the formula Z 1/p min Φ = N 3 − lim (N 3 − Φ)p dq p→∞ TN where the torus TN is endowed with its usual probability measure. Proof. This follows from the above formula, with Θ = N 3 − Φ. Observe that Θ is indeed 3 positive, because Φ is by definition a sum of N complex numbers of modulus 1. Let us restrict now the attention to the problem of computing the moments of Φ, which is more or less the same as computing those of N 3 − Φ. We have here: Proposition 6.25. The moments of Φ are given by Z p i1k1 . . . ipkp Φ dq = # is + ks = js + ls, [i1k1 . . . ipkp] = [j1l1 . . . jplp] N j1l1 . . . jplp T where the sets between brackets are by definition sets with repetition. Proof. This is indeed clear from the formula of Φ. See [19]. Regarding now the real case, an analogue of Proposition 6.25 holds, but the combina- torics does not get any simpler. One idea in dealing with this problem is by considering the “enveloping sum”, obtained from Φ by dropping the condition i + k = j + l: X qiqk Φ˜ = qjql ijkl The point is that the moments of Φ appear as “sub-quantities” of the moments of Φ,˜ so perhaps the question to start with is to understand very well the moments of Φ.˜ ˜ P 4 And this latter problem sounds like a quite familiar one, because Φ = | i qi| . We will be back to this a bit later. For the moment, let us do some combinatorics: Proposition 6.26. We have the moment formula Z X 2p N! Φ˜ p dq = N π (N − |π|)! T π∈P (2p) 2p 2p where = , with b1, . . . , b|π| being the lengths of the blocks of π. π b1,...,b|π| Proof. Indeed, by using the same method as for Φ, we obtain: Z ˜ p i1k1 . . . ipkp Φ(q) dq = # [i1k1 . . . ipkp] = [j1l1 . . . jplp] N j1l1 . . . jplp T The sets with repetitions on the right are best counted by introducing the corresponding partitions π = ker i1k1 . . . ipkp , and this gives the formula in the statement. 100 TEO BANICA In order to discuss now the real case, we have to slightly generalize the above result, by computing all the half-moments of Φ.e The result here is best formulated as: Proposition 6.27. We have the moment formula Z X 2p X N! | qi| dq = Cpk N (N − k)! T i k P p where Cpk = , with b1, . . . , b|π| being the lengths of the blocks of π. π∈P (p),|π|=k b1,...,b|π| Proof. This follows indeed exactly as Proposition 6.26 above, by replacing the exponent p by the exponent p/2, and by splitting the resulting sum as in the statement. Observe that the above formula basically gives the moments of Φ,˜ in the real case. Indeed, let us restrict attention to the case N = 2m. Then, for the purposes of our minimization problem we can assume that our vector is of the following form: q = (1, q1, . . . , qm−1, 1, q¯m−1,..., q¯1) So, we are led to the following conclusion, relating the real and complex cases: m−1 Proposition 6.28. Consider the variable X = q1 + ... + qm−1 over the torus T . (1) For the complex problem at N = m − 1, we have Φe = |X|4 (2) For the real problem at N = 2m, we have Φe = |2 + X + X¯|4. Proof. This is indeed clear from the definition of the enveloping sum Φ.e Finally, here is a random walk formulation of the problem: Proposition 6.29. The moments of Φ have the following interpretation: R p (1) First, the moments of the enveloping sum Φe count the loops of length 4p on the standard lattice ZN ⊂ RN , based at the origin. (2) R Φp counts those loops which are “piecewise balanced”, in the sense that each of the p consecutive 4-paths forming the loop satisfy i + k = j + l modulo N. Proof. The first assertion follows from the formula in the proof of Proposition 6.27, and the second assertion follows from the formula in Proposition 6.25. This statement looks quite encouraging, but passing from (1) to (2) is quite a delicate task, because in order to interpret the condition i + k = j + l we have to label the N coordinate axes of R by elements of the cyclic group ZN , and this is a quite unfamiliar operation. In addition, in the real case the combinatorics becomes more complex due to the symmetries of the parameter space, and no further results are available so far. HADAMARD MATRICES 101 7. Bistochastic matrices In this section and the next two ones we discuss certain analytic aspects of the complex Hadamard matrices, based on the various inequalities obtained in section 1. We will extend these inequalities to the complex case, and discuss them in detail. As a first, fundamental result, we have: Theorem 7.1. The complex Hadamard matrices, which form the manifold √ XN = MN (T) ∩ NUN can be analytically detected in two ways, as follows: √ N/2 (1) Given H ∈ √MN (T) we have | det(H)| ≤ N , with equality when H ∈ NUN . 2 (2) Given H ∈ NUN we have ||H||1 ≤ N , with equality when H ∈ MN (T). Proof. This is something that we already know in the real case, from Theorem 1.17 and Theorem 1.18 above, and the proof in the general case is similar: N (1) This follows indeed as in the real case, because if we denote by H1,...,HN ∈ T the rows of H, then we have, according to the definition of the determinant: | det(H)| = vol < H1,...,HN > ≤ ||H || × ... × ||H || √ 1 N = ( N)N The equality holds when H1,...,HN are pairwise orthogonal, as claimed. (2) This is something that we discussed in much detail in Proposition 6.17, Theorem 6.18 and Proposition 6.19, and which follows from Cauchy-Schwarz: !1/2 X X 2 2 ||H||1 = |Hij| ≤ N |Hij| = N ij ij The equality case holds when |Hij| = 1 for any i, j, as claimed. We will further discuss all this in section 9 below, with a few comments on (1), and with a detailed study of the condition (2), which is something quite fruitful. Regarding now the third and last basic inequality from the real case, namely the excess estimate from Theorem 1.19, this is something of a different nature, that we will discuss in this section, and in the next one. Let us begin with the following definition: Definition 7.2. A complex Hadamard matrix H ∈ MN (C) is called bistochastic when the sums on all rows and all columns are equal. We denote by n o bis XN = H ∈ XN H = bistochastic the real algebraic manifold formed by such matrices. 102 TEO BANICA The bistochastic Hadamard matrices are quite interesting objects, and include for in- stance all the circulant Hadamard matrices, that we discussed in section 6. Indeed, assuming that Hij = ξj−i is circulant, all rows and columns sum up to: X λ = ξi i So, let us first review the material in section 6, from this perspective. As a first and trivial remark, the Fourier matrix F2 looks better in bistochastic form: 1 1 i 1 F = ∼ = F 0 2 1 −1 1 i 2 This is something quite interesting, philosophically speaking. Indeed, we have here a new idea, namely that of studying the Hadamard matrices H ∈ MN (±1) by putting them 0 in complex bistochastic form, H ∈ MN (T), and then studying these latter matrices. We will see later on that, while certainly being viable, this idea is quite difficult to develop in practice, and is in need of some considerable preliminary work. Let us keep now reviewing the material in section 6. According to the results there, and to the above-mentioned fact that circulant implies bistochastic, we have: Theorem 7.3. The class of bistochastic Hadamard matrices is stable under permuting rows and columns, and under taking tensor products. As examples, we have: 0 (1) The circulant and symmetric forms FN of the Fourier matrices FN . 0 (2) The bistochastic and symmetric forms FG of the Fourier matrices FG. (3) The circulant and symmetric Backelin matrices, having size MN with M|N. Proof. In this statement the claim regarding permutations of rows and columns is clear. Assuming now that H,K are bistochastic, with sums λ, µ, we have: X X X X (H ⊗ K)ia,jb = HijKab = Hij Kab = λµ ia ia i a X X X X (H ⊗ K)ia,jb = HijKab = Hij Kab = λµ jb jb j b Thus, the matrix H ⊗K is bistochastic as well. As for the assertions (1,2,3), we already know all this, from Theorem 6.6, Theorem 6.8 and Theorem 6.9 above. In the above list of examples, (2) is the key entry. Indeed, while many interesting matrices, such as the usual Fourier ones FN , can be put in circulant form, this is something quite exceptional, which does not work any longer when looking at the general Fourier matrices FG. To be more precise, when using a decomposition of type G = ZN1 ×...×ZNk , and setting F 0 = F 0 ⊗ ... ⊗ F 0 , we can only say that F 0 is bistochastic. G N1 Nk G HADAMARD MATRICES 103 As a conclusion, the bistochastic Hadamard matrices are interesting objects, definitely worth some study. So, let us develop now some general theory, for such matrices. First, we have the following elementary result: Proposition 7.4. For an Hadamard matrix H ∈ MN (C), the following are equivalent: (1) H is bistochastic, with sums λ. (2) H is row-stochastic, with sums λ, and |λ|2 = N. Proof. Both the implications are elementary, as follows: N (1) =⇒ (2) If we denote by H1,...,HN ∈ T the rows of H, we have indeed: X N = < H1,H1 >= < H1,Hi > i X X ¯ X X ¯ = H1jHij = H1j Hij i j j i X ¯ ¯ 2 = H1j · λ = λ · λ = |λ| j N (2) =⇒ (1) Consider the all-one vector ξ = (1)i ∈ C . The fact that H is row- stochastic with sums λ, respectively column-stochastic with sums λ, reads: X X Hij = λ, ∀i ⇐⇒ Hijξj = λξi, ∀i ⇐⇒ Hξ = λξ j j X X t Hij = λ, ∀j ⇐⇒ Hijξi = λξj, ∀j ⇐⇒ H ξ = λξ i j We must prove that the first condition implies the second one, provided that the row sum λ satisfies |λ|2 = N. But this follows from the following computation: Hξ = λξ =⇒ H∗Hξ = λH∗ξ =⇒ N 2ξ = λH∗ξ =⇒ N 2ξ = λH¯ tξ =⇒ Htξ = λξ Thus, we have proved both the implications, and we are done. Here is another basic result, that we will need as well in what follows: Proposition 7.5. For a complex Hadamard matrix H ∈ MN (C), and a number λ ∈ C satisfying |λ|2 = N, the following are equivalent: (1) We have H ∼ H0, with H0 being bistochastic, with sums λ. N (2) Kij = aibjHij is bistochastic with sums λ, for some a, b ∈ T . (3) The equation Hb = λa¯ has solutions a, b ∈ TN . 104 TEO BANICA Proof. Once again, this is an elementary result, the proof being as follows: (1) ⇐⇒ (2) Since the permutations of the rows and columns preserve the bistochas- ticity condition, the equivalence H ∼ H0 that we are looking for can be assumed to come only from multiplying the rows and columns by numbers in T. Thus, we are looking for scalars ai, bj ∈ T such that Kij = aibjHij is bistochastic with sums λ, as claimed. (2) ⇐⇒ (3) The row sums of the matrix Kij = aibjHij are given by: X X Kij = aibjHij = ai(Hb)i j j Thus K is row-stochastic with sums λ precisely when Hb = λa¯, and by using the equivalence in Proposition 7.4, we obtain the result. Finally, here is an extension of the excess inequality from Theorem 1.19 above: Theorem 7.6. For a complex Hadamard matrix H ∈ MN (C), the excess, X E(H) = Hij ij √ satisfies |E(H)| ≤ N N, with equality if and only if H is bistochastic. N Proof. In terms of the all-one vector ξ = (1)i ∈ C , we have: X X ¯ X ¯ E(H) = Hij = Hijξjξi = (Hξ)iξi =< Hξ, ξ > ij ij i √ Now by using the Cauchy-Schwarz inequality, along with the fact that U = H/ N is unitary, and hence of norm 1, we obtain, as claimed: √ |E(H)| ≤ ||Hξ|| · ||ξ|| ≤ ||H|| · ||ξ||2 = N N Regarding now the equality case, this requires the vectors Hξ, ξ to be proportional, and so our matrix H to be row-stochastic. Moreover, if we assume Hξ = λξ, the above 2 computation gives |λ| = N, and by Proposition 7.4, we obtain the result. The above estimate is potentially quite useful, because it allows us to analytically locate bis the bistochastic Hadamard manifold XN inside the whole Hadamard manifold XN , a bit in the spirit of the two analytic methods in Theorem 7.1. We will be back to this later, with a number of probabilistic results on the subject. Let us go back now to the fundamental question, which already appeared several times in the above, of putting an arbitrary Hadamard matrix in bistochastic form. As already explained, we are interested in solving this question in general, and in particular in the real case, with potential complex reformulations of the HC and CHC at stake. What we know so far on this subject can be summarized as follows: HADAMARD MATRICES 105 Proposition 7.7. An Hadamard matrix H ∈ MN (C) can be put in bistochastic form when one of the following conditions is satisfied: √ N (1) The equations |Ha|i = N, with i = 1,...,N√ , have solutions a ∈ T . (2) The quantity |E| attains its maximum N N over the equivalence class of H. Proof. This follows indeed from Proposition 7.4 and Proposition 7.5 above. Thus, we have two approaches to the problem, one algebraic, and one analytic. Let us first discuss the algebraic approach, coming from (1) above. What we have there is a certain system of N equations, having as unknowns N real variables, namely the phases of a1, . . . , aN . This system is highly non-linear, but can be solved, however, via a certain non-explicit method, as explained by Idel and Wolf in [57]. In order to discuss this material, which is quite advanced, let us begin with some preliminaries. The complex projective space appears by definition as follows: P N−1 = ( N − {0}) < x = λy > C C Inside this projective space, we have the Clifford torus, constructed as follows: n o N−1 = (z , . . . , z ) ∈ P N−1 |z | = ... = |z | T 1 N C 1 N With these conventions, we have the following result, from [57]: Proposition 7.8. For a unitary matrix U ∈ UN , the following are equivalent: 0 (1) There exist L, R ∈ UN diagonal such that U = LUR is bistochastic. (2) The standard torus TN ⊂ CN satisfies TN ∩ UTN 6= ∅. (3) The Clifford torus N−1 ⊂ P N−1 satisfies N−1 ∩ U N−1 6= ∅. T C T T Proof. These equivalences are all elementary, as follows: (1) =⇒ (2) Assuming that U 0 = LUR is bistochastic, which in terms of the all-1 vector ξ means U 0ξ = ξ, if we set f = Rξ ∈ TN we have: 0 0 N Uf = LU¯ Rf¯ = LU¯ ξ = Lξ¯ ∈ T Thus we have Uf ∈ TN ∩ UTN , which gives the conclusion. (2) =⇒ (1) Given g ∈ TN ∩ UTN , we can define R,L as follows: ¯ R = diag(g1, . . . , gN ) , L = diag((Ug)1,..., (Ug)N ) We have then Rξ = g and Lξ¯ = Ug, and so U 0 = LUR is bistochastic, because: U 0ξ = LURξ = LUg = ξ (2) =⇒ (3) This is clear, because N−1 ⊂ P N−1 appears as the projective image of T C TN ⊂ CN , and so TN−1 ∩ UTN−1 appears as the projective image of TN ∩ UTN . (3) =⇒ (2) We have indeed the following equivalence: N−1 N−1 N N T ∩ UT 6= ∅ ⇐⇒ ∃λ 6= 0, λT ∩ UT 6= ∅ But U ∈ UN implies |λ| = 1, and this gives the result. 106 TEO BANICA The point now is that the condition (3) above is something familiar in symplectic geometry, and known to hold for any U ∈ UN . Thus, following [57], we have: Theorem 7.9. Any unitary matrix U ∈ UN can be put in bistochastic form, U 0 = LUR with L, R ∈ UN being both diagonal, via a certain non-explicit method. Proof. As already mentioned, the condition TN−1 ∩ UTN−1 6= ∅ in Proposition 7.8 (3) is something quite natural in symplectic geometry. To be more precise, N−1 ⊂ P N−1 is T C a Lagrangian submanifold, TN−1 → UTN−1 is a Hamiltonian isotopy, and a result from [29], [35] states that TN−1 cannot be displaced from itself via a Hamiltonian isotopy. Thus, the results in [29], [35] tells us that TN−1 ∩ UTN−1 6= ∅ holds indeed, for any U ∈ UN . We therefore obtain the result, via Proposition 7.8. See [57]. In relation now with our Hadamard matrix questions, we have: Theorem 7.10. Any complex Hadamard matrix can be put in bistochastic form, up to the standard equivalence relations for such matrices. √ Proof. This√ follows indeed from Theorem 7.9, because if H = NU is Hadamard then so is H0 = NU 0, and with the remark that, in what regards the equivalence relation, we just need the multiplication of the rows and columns by scalars in T. All this is extremely interesting, but unfortunately, not explicit. As explained in [57], the various technical results from [29], [35] show that in the generic, “transverse” situation, N−1 there are at least 2 ways of putting a unitary matrix U ∈ UN in bistochastic form, and this modulo the obvious transformation U → zU, with |z| = 1. Summarizing, the question of explicitely putting the Hadamard matrices H ∈ MN (C) in bistochastic form remains open, and open as well is the question of finding a simpler proof for the fact that this can be done indeed, without using [29], [35]. Regarding this latter question, a possible approach comes from the excess result from Theorem 7.6 above. Indeed, in view of the remark√ there, it is enough to show that the law of |E| over the equivalence class of H has N N as upper support bound. In order to comment on this, let us first formulate: Definition 7.11. The glow of H ∈ MN (C) is the measure µ ∈ P(C) given by: ! Z Z X ϕ(x)dµ(x) = ϕ aibjHij d(a, b) N N C T ×T ij P That is, the glow is the law of E = ij Hij, over the equivalence class of H. HADAMARD MATRICES 107 In this definition H can be any complex matrix, but the equivalence relation is the one for the complex Hadamard matrices. To be more precise, let us call two matrices H,K ∈ MN (C) equivalent if one can pass from one to the other by permuting rows and columns, or by multiplying the rows and columns by numbers in T. Now since P permuting rows and columns does not change the quantity E = ij Hij, we can restrict N N attention from the full equivalence group G = (SN o T ) × (SN o T ) to the smaller group G0 = TN × TN , and we obtain the measure µ in Definition 7.11. As in the real case, the terminology comes from a picture of the following type, with the stars ∗ representing the entries of our matrix, and with the switches being supposed now to be continuous, randomly changing the phases of the concerned entries: → ∗ ∗ ∗ ∗ → ∗ ∗ ∗ ∗ → ∗ ∗ ∗ ∗ → ∗ ∗ ∗ ∗ ↑ ↑ ↑ ↑ In short, what we have here is a complex generalization of the Gale-Berlekamp game [50], [82], and this is where the main motivation for studying the glow comes from. We are in fact interested in computing a real measure, because we have: Proposition 7.12. The laws µ, µ+ of E, |E| over the torus TN × TN are related by µ = ε × µ+ where × is the multiplicative convolution, and ε is the uniform measure on T. Proof. We have E(λH) = λE(H) for any λ ∈ T, and so µ = law(E) is invariant under the action of T. Thus µ must decompose as µ = ε × µ+, where µ+ is a certain probability + measure on [0, ∞), and this measure µ is the measure in the statement. In particular, we can see from the above result that the glow is invariant under rotations. With this observation made, we can formulate the following result: Theorem√ 7.13. The glow of any Hadamard matrix H ∈ MN (C), or more generally of any H ∈ NUN , satisfies the following conditions, where D is the unit disk, √ √ N N T ⊂ supp(µ) ⊂ N N D with the inclusion on the right coming from Cauchy-Schwarz, and with the inclusion on the left corresponding to the fact that H can be put in bistochastic form. 108 TEO BANICA Proof. In this statement the inclusion on the right comes indeed from Cauchy-Schwarz, as explained in the proof of Theorem 7.6 above, with the√ remark that the computation there only uses the fact that the rescaled matrix U = H/ N is unitary. Regarding now the inclusion on the left, we know from Theorem 7.9 that H can be put in bistochastic form. According to Proposition 7.7, this tells us that we have: √ N N T ∩ supp(µ) 6= ∅ Now by using the rotational invariance of the√ glow, and hence of its support, coming from Proposition 7.12, we obtain from this N N T ⊂ supp(µ), as claimed. The challenging question is that of obtaining a proof of the above result by using probabilistic techniques. Indeed, as explained at the end of section 6 above, the support of a measure can be recaptured from the moments, simply by computing a limit. Thus, knowing the moments of the glow well enough would solve the problem. Regarding the moments of the glow, the formula is as follows: Proposition 7.14. For H ∈ MN (T) the even moments of |E| are given by Z X Hi1j1 ...Hipjp |E|2p = N N Hk l ...Hk l T ×T [i]=[k],[j]=[l] 1 1 p p where the sets between brackets are by definition sets with repetition. Proof. We have indeed the following computation: Z Z 2p 2p X |E| = Hijaibj N N N N T ×T T ×T ij Z !p X Hij aibj = · N N Hkl akbl T ×T ijkl Z Z X Hi1j1 ...Hipjp ai1 . . . aip bj1 . . . bjp = Hk l ...Hk l N ak . . . ak N bl . . . bl ijkl 1 1 p p T 1 p T 1 p Now since the integrals at right equal respectively the Kronecker symbols δ[i],[k] and δ[j],[l], we are led to the formula in the statement. With this formula in hand, the main result, regarding the fact that the complex Hada- mard matrices can be put in bistochastic form, reformulates as follows: Theorem 7.15. For a complex Hadamard matrix H ∈ MN (T) we have 1/p X Hi1j1 ...Hipjp 3 lim = N p→∞ Hk l ...Hk l [i]=[k],[j]=[l] 1 1 p p coming from the fact that H can be put in bistochastic form. HADAMARD MATRICES 109 Proof. This follows from the well-known fact that the maximum of a bounded function Θ: X → [0, ∞) can be recaptured via following formula: Z 1/p max(Θ) = lim Θ(x)p dx p→∞ X With X = TN × TN and with Θ = |E|2, we conclude that the limit in the statement is the square of the upper bound of the glow. But, according to Theorem 7.13, this upper 3 bound is known to be ≤ N by Cauchy-Schwarz, and the equality holds by [57]. To conclude now, the challenging question is that of finding a direct proof for Theorem 7.15. All this would provide an alternative aproach to the results in [57], which would be of course still not explicit, but which would use at least some more familiar tools. We will discuss such questions in section 9 below, with the remark however that the problems at N ∈ N fixed being quite difficult, we will do a N → ∞ study only. Getting away now from these difficult questions, we have nothing concrete so far, besides the list of examples from Theorem 7.3, coming from the circulant matrix considerations in section 6. So, our purpose will be that of extending that list. A first natural question is that of looking at the Butson matrix case. We have here the following result, extending the finding from Proposition 7.4 above: Proposition 7.16. Assuming that HN (l) contains a bistochastic matrix, the equations a0 + a1 + ... + al−1 = N l−1 2 |a0 + a1w + ... + al−1w | = N must have solutions, over the positive integers. Proof. This is a reformulation of the equality |λ|2 = N, from Proposition 7.4 above. 2πi/l i Indeed, if we set w = e , and we denote by ai ∈ N the number of w entries appearing in the first row of our matrix, then the row sum of the matrix is given by: l−1 λ = a0 + a1w + ... + al−1w Thus, we obtain the system of equations in the statement. The point now is that, in practice, we are led precisely to the Turyn obstructions from section 6 above. At very small values of l, the obstructions are as follows: Theorem 7.17. Assuming that HN (l) contains a bistochastic matrix, the following equa- tions must have solutions, over the integers: (1) l = 2: 4n2 = N. (2) l = 3: x2 + y2 + z2 = 2N, with x + y + z = 0. (3) l = 4: a2 + b2 = N. 110 TEO BANICA Proof. This follows indeed from the results that we have: (1) This is something well-known, which follows from Proposition 7.16. (2) This is best viewed by using Proposition 7.16, and the following formula, that we already know, from section 3 above: 1 |a + bw + cw2|2 = [(a − b)2 + (b − c)2 + (c − a)2] 2 At the level of the concrete obstructions, we must have for instance 56 |N. Indeed, this follows as in the proof of the de Launey obstruction for HN (3) with 5|N. 2 2 2 (3) This follows again from Proposition 7.16, and from |a + ib| = a + b . As a conclusion, nothing much interesting is going on in the Butson matrix case, with various arithmetic obstructions, that we partly already met, appearing here. See [63]. In order to reach, however, to a number of positive results, beyond those in Theorem 7.3, we can investigate various special classes of matrices, such as the Dit¸˘aproducts. In order to formulate our results, we will use the following notion: Definition 7.18. We say that a complex Hadamard√ matrix H ∈ MN (C) is in “almost bistochastic form” when all the row sums belong to N · T. Observe that, assuming that this condition holds, the matrix H can be put in bistochas- tic form, just by multiplying its rows by suitable numbers from T. We will be particularly interested here in the special situation where the affine defor- q mations H ∈ MN (C) of a given complex Hadamard matrix H ∈ MN (C) can be put in almost bistochastic form, independently of the value of the parameter q. For the simplest deformations, namely those of F2 ⊗ F2, this is indeed the case: p q Proposition 7.19. The deformations of F2 ⊗ F2, with parameter matrix Q = (r s), p q p q p −q p −q F2 ⊗Q F2 = r s −r −s r −s −r s can be put in almost bistochastic form, independently of the value of Q. Proof. By multiplying the columns of the matrix in the statement with 1, 1, −1, 1 respec- tively, we obtain the following matrix: p q −p q 00 p −q −p −q F2 ⊗ F2 = Q r s r −s r −s r s The row sums of this matrix being 2q, −2q, 2r, 2r ∈ 2T, we are done. HADAMARD MATRICES 111 00 We will see later on that the above matrix F2 ⊗Q F2 is equivalent to a certain matrix 0 0 F2 ⊗ F2, which looks a bit more complicated, but is part of a series FN ⊗ FN . Now back to the general case, we have the following result: Theorem 7.20. A deformed tensor product H ⊗Q K can be put in bistochastic form when i there exist numbers xa ∈ T such that with ∗ i (K x )b Gib = Qib √ ∗ we have |(H G)ib| = MN, for any i, b. Proof. The deformed tensor product L = H ⊗Q K is given by Lia,jb = QibHijKab. By multiplying the columns by scalars Rjb ∈ T, this matrix becomes: 0 Lia,jb = RjbQibHijKab The row sums of this matrix are given by: 0 X Sia = RjbQibHijKab jb X X = KabQib HijRjb b j X = KabQib(HR)ib b i In terms of the variables Cb = Qib(HR)ib, these rows sums are given by: 0 X i i Sia = KabCb = (KC )a b Thus H ⊗Q K can be put in bistochastic form when we can find scalars Rjb ∈ T and i i xa ∈ T such that, with Cb = Qib(HR)ib, the following condition is satisfied: √ i i (KC )a = MNxa , ∀i, a √ But this condition is equivalent to KCi = MNxi for any i, and by multiplying to the left by the adjoint matrix K∗, we are led to the following condition: √ √ NCi = MK∗xi , ∀i i Now by recalling that Cb = Qib(HR)ib, this condition is equivalent to: √ √ ∗ i NQib(HR)ib = M(K x )b , ∀i, b ∗ i With Gib = (K x )b/Qib as in the statement, this condition reads: √ √ N(HR)ib = MGib , ∀i, b 112 TEO BANICA √ √ But this condition is equivalent to NHR = MG, and by multiplying to the left by the adjoint matrix H∗, we are led to the following condition: √ MNR = H∗G Thus, we have obtained the condition in the statement. As an illustration for this result, assume that H,K can be put in bistochastic form, by M N i using vectors y ∈ T , z ∈ T . If we set xa = yiza, with Q = 1 we have: ∗ i ∗ ∗ Gib = (K x )b = [K (yiz)]b = yi(K z)b We therefore obtain the following formula: ∗ X ∗ (H G)ib = (H )ijGjb j X ∗ ∗ = (H )ijyj(K z)b j ∗ ∗ = (H y)i(K z)b Thus the usual tensor product H ⊗ K can be put in bistochastic form as well. In the case H = FM the equations simplify, and we have: Proposition 7.21. A deformed tensor product FM ⊗Q K can be put in bistochastic form i when there exist numbers xa ∈ T such that with ∗ i (K x )b Gib = Qib we have the following formulae, with l being taken modulo M: X ¯ GjbGj+l,b = MNδl,0 , ∀l, b j 2 2 Moreover, the M×N matrix |Gjb| is row-stochastic with sums N , and the l = 0 equations state that this matrix must be column-stochastic, with sums MN. Proof. With notations from Theorem 7.20, and with w = e2πi/M , we have: ∗ X −ij (H G)ib = w Gjb j HADAMARD MATRICES 113 The absolute value of this number can be computed as follows: ∗ 2 X i(k−j) ¯ |(H G)ib| = w GjbGkb jk X il ¯ = w GjbGj+l,b jl X il X ¯ = w GjbGj+l,b l j b If we denote by vl the sum on the right, we obtain: ∗ 2 X il b b |(H G)ib| = w vl = (FM v )i l √ M ∗ Now if we denote by ξ the all-one vector in C , the condition |(H G)ib| = MN for any i, b found in Theorem 7.20 above reformulates as follows: F M vb = MNξ , ∀b ∗ By multiplying to the left by FM /M, this condition is equivalent to: MN 0 b ∗ v = NF ξ = . , ∀b M . 0 b b Let us examine the first equation, v0 = MN. By definition of vl , we have: b X ¯ X 2 v0 = GjbGjb = |Gjb| j j ∗ j Now recall from Theorem 7.20 that we√ have Gjb = (K x )b/Qjb, for certain numbers j ∗ xb ∈ T. Since we have Qjb ∈ T and K / N ∈ UN , we obtain: X 2 X ∗ j 2 |Gjb| = |(K x )b| b b ∗ j 2 = ||K x ||2 j 2 = N||x ||2 = N 2 2 2 Thus the M × N matrix |Gjb| is row-stochastic, with sums N , and our equations b v0 = MN for any b state that this matrix must be column-stochastic, with sums MN. 114 TEO BANICA b Regarding now the other equations that we found, namely vl = 0 for l 6= 0, by definition b of vl and of the variables Gjb, these state that we must have: X ¯ GjbGj+l,b = 0 , ∀l 6= 0, ∀b j Thus, we are led to the conditions in the statement. As an illustration for this result, let us go back to the Q = 1 situation, explained after ∗ Theorem 7.20. By using the formula Gib = yi(K z)b there, we have: X ¯ X ∗ ∗ GjbGj+l,b = yj(K z)b yj+l(K z)b j j X yj = |(K∗z) |2 b y j j+l = M · Nδl,0 Thus, if K can be put in bistochastic form, then so can be put FM ⊗ K. 0 As a second illustration, let us go back to the matrices F2 ⊗Q F2 from the proof of Proposition 7.19 above. The vector of the row sums being S = (2q, −2q, 2r, 2r), we have x = (q, −q, r, r), and so we obtain the following formulae for the entries of G: 1 1 q 0 1 −1 −q 2q b b G0b = = Q0b Q0b 1 1 r 2r 1 −1 r 0 b b G1b = = Q1b Q1b Thus, in this case the matrix G is as follows, independently of Q: 0 2 G = 2 0 In particular, we see that the conditions in Proposition 7.21 are satisfied. As a main application now, we have the following result: Theorem 7.22. The Dit¸˘adeformations FN ⊗Q FN can be put in almost bistochastic form, independently of the value of the parameter matrix Q ∈ MN (T). 2πi/N Proof. We use Proposition 7.21 above, with M = N, and with K = FN . Let w = e , and consider the vectors xi ∈ TN given by: i (i−1)a x = (w )a HADAMARD MATRICES 115 ∗ i Since K K = N1N , and x are the column vectors of K, shifted by 1, we have: 0 N 0 0 0 0 ∗ 0 . ∗ 1 . ∗ N−1 . K x = . ,K x = . ,...,K x = . 0 0 N N 0 0 ∗ i We conclude that we have (K x )b = Nδi−1,b, and so the matrix G is given by: Nδi−1,b Gib = Qib With this formula in hand, the sums in Proposition 7.21 are given by: X X Nδj−1,b Nδj+l−1,b G G¯ = · jb j+l,b Q Q j j jb j+l,b In the case l 6= 0 we clearly get 0, because the products of Kronecker symbols are 0. In 2 the case l = 0 the denominators are |Qjb| = 1, and we obtain: X ¯ 2 X 2 GjbGjb = N δj−1,b = N j j Thus, the conditions in Proposition 7.21 are satisfied, and we obtain the result. Here is an equivalent formulation of the above result: 0 Theorem 7.23. The matrix FN ⊗Q FN , with Q ∈ MN (T), defined by ij+ab 0 w Qib (FN ⊗Q FN )ia,jb = bj+j · w Qb+1,b 2πi/N where w = e is almost bistochastic, and equivalent to FN ⊗Q FN . Proof. Our claim is that this is the matrix constructed in the proof of Theorem 7.22. Indeed, let us first go back to the proof of Theorem 7.20. In the case M = N and H = K = FN , the Dit¸˘adeformation L = H ⊗Q K studied there is given by: ij+ab Lia,jb = QibHijKab = w Qib As explained in the proof of Theorem 7.22, if the conditions in the statement there are 0 satisfied, then the matrix Lia,jb = RjbLia,jb is almost bistochastic, where: √ MN · R = H∗G In our case now, M = N and H = K = FN , we know from the proof of Proposition 7.21 that the choice of G which makes work Theorem 7.22 is as follows: Nδi−1,b Gib = Qib 116 TEO BANICA With this formula in hand, we can compute the matrix R, as follows: 1 R = (H∗G) jb N jb 1 X = w−ijG N ib i X δi−1,b = wij · Q i ib w−(b+1)j = Qb+1,b Thus, the modified version of FN ⊗Q FN which is almost bistochastic is given by: 0 Lia,jb = RjbLia,jb −(b+1)j w ij+ab = · w Qib Qb+1,b ij+ab w Qib = bj+j · w Qb+1,b Thus we have obtained the formula in the statement, and we are done. As an illustration, let us work out the case N = 2. Here we have w = −1, and with p q p s Q = (r s), and then with u = r , v = q , we obtain the following matrix: p q p q r q − r q u 1 −u 1 p q p q r − q − r − q u −1 −u −1 F2 ⊗Q F2 = r s r s = r q r − q 1 v 1 −v r s r s 1 −v 1 v r − q r q Observe that this matrix is indeed almost bistochastic, with row sums 2, −2, 2, 2. It is quite unclear on how to get beyond these results. An interesting question here would be probably that of focusing on the real case, and see if the Hadamard matrices there, H ∈ MN (±1), can be put or not in bistochastic form, in an explicit way. This is certainly true for the Walsh matrices, but for the other basic examples, such as the Paley or the Williamson matrices, no results seem to be known so far. Having such a theory would be potentially very interesting, with a complex reformula- tion of the HC and of the other real Hadamard questions at stake. HADAMARD MATRICES 117 8. Glow computations We discuss here the computation of the glow, in the N → ∞ limit. We have seen in section 1 that, in what concerns the real glow of the real Hadamard matrices, with N → ∞ we obtain a real Gaussian measure. Our main purpose here will be that of establishing a similar result in the complex case, stating that for the complex glow of the complex Hadamard matrices, with N → ∞ we obtain a complex Gaussian measure. The computations in the complex case are considerably more involved than those in the real case, and will require a lot of combinatorics. Also, we will investigate the problem of getting beyond the N → ∞ limiting result, with formulae at order 1, 2, 3, 4. As a first motivation for all this, we have the Gale-Berlekamp game [50], [82]. Another motivation comes from the questions regarding the bistochastic matrices, in relation with [57], explained in section 7. Finally, we have the question of connecting and computing the defect, and other invariants of the Hadamard matrices, in terms of the glow. Let us begin by reviewing the few theoretical things that we know about the glow, from section 7 above. The main results there can be summarized as follows: Theorem 8.1. The glow of H ∈ MN (C), which is the law µ ∈ P(C) of the excess X E = Hij ij over the Hadamard equivalence class of H, has the following properties: (1) µ = ε × µ+, where µ+ = law(|E|). (2) µ is invariant√ under rotations. √ (3) H ∈ √NUN implies supp(µ) ⊂√N N D. (4) H ∈ NUN implies as well N N T ⊂ supp(µ). Proof. We already know all this from section 7, the idea being as follows: (1) This follows by using H → zH with |z| = 1, as explained in Proposition 7.12. (2) This follows from (1), the convolution with ε bringing the invariance. (3) This folllows from Cauchy-Schwarz, as explained in Theorem 7.13. (4) This is something highly non-trivial, coming from [57]. In what follows we will be mainly interested in the Hadamard matrix case, but since the computations here are quite difficult, let us begin our study with other matrices. It is convenient to normalize our matrices, by assuming that the corresponding 2-norm qP 2 ||H||2 = |Hij| takes the value ||H||2 = N. Note that this is always the case with ij √ the Hadamard matrices, and more generally with the matrices H ∈ NUN . We recall that the complex Gaussian distribution C is the law of z = √1 (x + iy), where 2 x, y are independent standard Gaussian variables. In order to detect this distribution, we can use the moment method, and the well-known formula E(|z|2p) = p!. 118 TEO BANICA Finally, we use the symbol ∼ to denote an equality of distributions. We have: Proposition 8.2. We have the following computations: √ √ N (1) For the rescaled identity IeN = NIN we have E ∼ N(q1 +...+qN ), with q ∈ T random. With N → ∞ we have E/N ∼ C. (2) For the flat matrix JN = (1)ij we have E ∼ (a1 + ... + aN )(b1 + ... + bN ), with (a, b) ∈ TN × TN random. With N → ∞ we have E/N ∼ C × C. Proof. We use Theorem√ 8.1, and the moment method: P N (1) Here we have E = N i aibi, with a, b ∈ T random, and with qi = aibi this gives the first assertion. Let us estimate now the moments of |E|2. We have: Z Z 2p p 2p |E| = N |q1 + ... + qN | dq TN ×TN TN Z X qi1 . . . qip = N p dq N q . . . q T ij j1 jp n o p p p = N # (i, j) ∈ {1,...,N} × {1,...,N} [i1, . . . , ip] = [j1, . . . , jp] ' N p · p!N(N − 1) ... (N − p + 1) ' N p · p!N p = p!N 2p Here, and in what follows, the sets between brackets are by defintion sets with repetition, and the middle estimate comes from the fact that, with N → ∞, only the multi-indices i = (i1, . . . , ip) having distinct entries contribute. But this gives the result. P P P (2) Here we have E = ij aibj = i ai j bj, and this gives the first assertion. Now N P P since a, b ∈ T are independent, so are the quantities i ai, j bj, so we have: Z Z 2 2p 2p p 2 |E| = |q1 + ... + qN | dq ' (p!N ) TN ×TN TN Here we have used the estimate in the proof of (1), and this gives the result. As a first conclusion, the glow is intimately related to the basic hypertoral law, namely N that of q1 + ... + qN , with q ∈ T random. Observe that at N = 1 this hypertoral law is simply δ1, and that at N = 2 we obtain the following law: law|1 + q| = lawp(1 + eit)(1 + e−it) √ = law 2 + 2 cos t t = law 2 cos 2 HADAMARD MATRICES 119 P In general, the law of qi is known to be related to the P´olya random walk [78]. Also, as explained for instance in section 6, the moments of this law are: Z 2p X p N! |q1 + ... + qN | dq = N π (N − |π|)! T π∈P (p) As a second conclusion, even under the normalization ||H||2 = N, the glow can behave quite differently in the N → ∞ limit. So, let us restrict now the attention to the Hadamard matrices. At N = 2 we only have F2 to be invesigated, the result being: Proposition 8.3. For the Fourier matrix F2 we have |E|2 = 4 + 2Re(α − β) for certain variables α, β ∈ T which are uniform, and independent. Proof. The matrix that we interested in, namely the Fourier matrix F2 altered by a vertical switching vector (a, b) and an horizontal switching vector (c, d), is: ac ad F = e2 bc −bd With this notation, we have the following formula: ad bc bd ac |E|2 = |ac + ad + bc − bd|2 = 4 + + − − bc ad ac bd ad bd For proving that α = bc and β = ac are independent, we use the moment method: Z p q Z Z Z Z ad bd p−q q−p −p−q p+q = a b c d = δp,q,0 T4 bc ac T T T T Thus α, β are indeed independent, and we are done. Observe that law(|E|2), and hence law(E), is uniquely determined by the above result. It is possible of course to derive from this some more concrete formulae, but let us look instead at the case N = 3. Here the matrix that we are interested in is: ad ae af 2 Fe3 = bd wbe w bf cd w2ce wcf Thus, we would like to compute the law of the following quantity: |E| = |ad + ae + af + bd + wbe + w2bf + cd + w2ce + wcf| 120 TEO BANICA The problem is that when trying to compute |E|2, the terms won’t cancel much. More 2 2 precisely, we have |E| = 9 + C0 + C1w + C2w , where C0,C1,C2 are as follows: ae ae af af bd bd be bf cd cd ce cf C = + + + + + + + + + + + 0 bd cd bd cd ae af cf ce ae af bf be ad ad ae af bd be be be cd cf cf cf C = + + + + + + + + + + + 1 bf ce bf ce ce ad af cd bf ad ae bd ad ad ae af bd bf bf bf cd ce ce ce C = + + + + + + + + + + + 2 be cf cf be cf ad ae cd be ad af bd In short, all this leads nowhere, and the exact study stops at F2. In general now, one idea is that of using Bernoulli-type variables coming from the row sums, as in the real case. We have here the following result: Theorem 8.4. The glow of H ∈ MN (C) is given by the formula Z law(E) = B((Ha)1,..., (Ha)N ) a∈TN P N where B(c1, . . . , cN ) = law( i λici), with λ ∈ T random. Proof. This is indeed clear from the formula E =< a, Hb >, because when a ∈ TN is fixed, E follows the law B((Ha)1,..., (Ha)N ) in the statement. Observe that we can write B(c1, . . . , cN ) = ε × β(|c1|,..., |cN |), where the measure P β(r1, . . . , rN ) ∈ P(R+) with r1, . . . , rN ≥ 0 is given by β(r1, . . . , rN ) = law| i λiri|. Regarding now the computation of β, we have: s X λi β(r , . . . , r ) = law · r r 1 N λ i j ij j Consider now the following variable, which is easily seen, for instance by using the moment method, to be uniform over the projective torus TN−1 = TN /T: λ1 λ2 λN (µ1, µ2, . . . , µN ) = , ,..., λ2 λ3 λ1 Now since we have λi/λj = µiµi+1 . . . µj, with the convention µi . . . µj = µj . . . µi for i > j, this gives the following formula, with µ ∈ TN−1 random: sX β(r1, . . . , rN ) = law µiµi+1 . . . µj · rirj ij It is possible to further study the laws β by using this formula. However, in practice, it is more convenient to use the complex measures B from Theorem 8.4. HADAMARD MATRICES 121 Let us end these preliminaries with a discussion of the “arithmetic” version of the problem, which makes the link with the Gale-Berlekamp switching game [50], [82] and with the work in section 1. We have the following unifying formalism: Definition 8.5. Given H ∈ MN (C) and s ∈ N ∪ {∞}, we define µs ∈ P(C) by ! Z Z X ϕ(x)dµs(x) = ϕ aibjHij d(a, b) N N C Zs ×Zs ij where Zs ⊂ T is the group of the s-roots of unity, with the convention Z∞ = T. Observe that at s = ∞ we obtain the measure in Theorem 8.1. Also, at s = 2 and for a usual Hadamard matrix, H ∈ MN (±1), we obtain the measure from section 1. Observe that for H ∈ MN (±1), knowing µ2 is the same as knowing the statistics of the number of one entries, |1 ∈ H|. This follows indeed from the following formula: X 2 Hij = |1 ∈ H| − | − 1 ∈ H| = 2|1 ∈ H| − N ij More generally, at s = p prime, we have the following result: Theorem 8.6. When s is prime and H ∈ MN (Zs), the statistics of the number of one P entries, |1 ∈ H|, can be recovered from that of the total sum, E = ij Hij. n Proof. The problem here is of vectorial nature, so given V ∈ Zs , we would like to compare P the quantities |1 ∈ V | and Vi. Let us write, up to permutations: V = (1 ... 1 w . . . w ...... ws−1 . . . ws−1) | {z } | {z } | {z } a0 a1 as−1 P s−1 We have then |1 ∈ V | = a0 and Vi = a0 + a1w + ... + as−1w , and we also know that a0 + a1 + ... + as−1 = n. Now when s is prime, the only ambiguity in recovering a0 s−1 s−1 from a0 + a1w + ... + as−1w can come from 1 + w + ... + w = 0. But since the sum of the numbers ai is fixed, a0 + a1 + ... + as−1 = n, this ambiguity dissapears. As an illustration, at s = 3 we can write E = a + bw + cw2, and we have: b + c n − a 3a − n Re(E) = a − = a − = 2 2 2 At s = 5, however, is it now clear how to explicitely solve the problem. In fact, the problem of finding an explicit formula is related to the question of relating the complex probability measures µs = law(E) for a given matrix H ∈ MN (T), at various values of s ∈ N ∪ {∞}. Note that Theorem 8.6 tells us that for a matrix H ∈ MN (Zs) with s prime, the measure µ2 can be recaptured from the knowledge of µs. Let us investigate now the glow of the complex Hadamard matrices, by using the moment method. We use the moment formula from section 7, namely: 122 TEO BANICA Proposition 8.7. For H ∈ MN (T) the even moments of |E| are given by Z X Hi1j1 ...Hipjp |E|2p = N N Hk l ...Hk l T ×T [i]=[k],[j]=[l] 1 1 p p where the sets between brackets are by definition sets with repetition. P Proof. As explained in section 7, with E = ij Hijaibj we obtain: Z Z !p X Hij aibj |E|2p = · N N N N Hkl akbl T ×T T ×T ijkl Z Z X Hi1j1 ...Hipjp ai1 . . . aip bj1 . . . bjp = Hk l ...Hk l N ak . . . ak N bl . . . bl ijkl 1 1 p p T 1 p T 1 p The integrals on the right being δ[i],[k] and δ[j],[l], we obtain the result. As a first application, let us investigate the tensor products. We have: Proposition 8.8. The even moments of |E| for L = H ⊗ K are given by Z X Hi1j1 ...Hipjp Ka1b1 ...Kapbp |E|2p = · NM NM Hk l ...Hk l Kc d ...Kc d T ×T [ia]=[kc],[jb]=[ld] 1 1 p p 1 1 p p where the sets between brackets are as usual sets with repetition. Proof. With L = H ⊗ K, the formula in Proposition 8.7 reads: Z X Li1a1,j1b1 ...Lipap,jpbp |E|2p = NM NM Lk c ,l d ...Lk c ,l d T ×T [ia]=[kc],[jb]=[ld] 1 1 1 1 p p p p But this gives the formula in the statement, and we are done. The above result is quite bad news. Indeed, we cannot reconstruct the glow of H ⊗ K from that of H,K, because the indices “get mixed”. Let us develop now some moment machinery. Let P (p) be the set of partitions of {1, . . . , p}, with its standard order relation ≤, which is such that uu ... ≤ π ≤ | | ... |, for any π ∈ P (p). We denote by µ(π, σ) the associated M¨obius function, given by: 1 if π = σ P µ(π, σ) = − π≤τ<σ µ(π, τ) if π < σ 0 if π 6≤ σ HADAMARD MATRICES 123